Text-to-Image
Diffusion Single File
qwen-image
qwen-image-2.1
nvfp4
comfyui
quantized
image-editing
rtx-5090
Instructions to use BennyDaBall/Qwen-Image-2.1-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusion Single File
How to use BennyDaBall/Qwen-Image-2.1-NVFP4 with Diffusion Single File:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Add tested native workflows, comparisons, encoder patch and validation
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +86 -0
- COMPARISONS.md +34 -0
- PROMPTING.md +71 -0
- README.md +132 -0
- SHA256SUMS +126 -0
- VALIDATION.md +125 -0
- assets/01_Text_to_Image.png +3 -0
- assets/02_Image_Editing.png +3 -0
- assets/03_Transparent_RGBA.png +3 -0
- assets/04_2K_Typography.png +3 -0
- assets/ab_architecture_20260920.jpg +3 -0
- assets/ab_architecture_20260921.jpg +3 -0
- assets/ab_astronaut_20260920.jpg +3 -0
- assets/ab_astronaut_20260921.jpg +3 -0
- assets/ab_botanical_poster_20260920.jpg +3 -0
- assets/ab_botanical_poster_20260921.jpg +3 -0
- assets/ab_fashion_20260920.jpg +3 -0
- assets/ab_fashion_20260921.jpg +3 -0
- assets/ab_illustration_20260920.jpg +3 -0
- assets/ab_illustration_20260921.jpg +3 -0
- assets/ab_landscape_20260920.jpg +3 -0
- assets/ab_landscape_20260921.jpg +3 -0
- assets/ab_macro_20260920.jpg +3 -0
- assets/ab_macro_20260921.jpg +3 -0
- assets/ab_night_market_20260920.jpg +3 -0
- assets/ab_night_market_20260921.jpg +3 -0
- assets/ab_portrait_20260920.jpg +3 -0
- assets/ab_portrait_20260921.jpg +3 -0
- assets/ab_product_20260920.jpg +3 -0
- assets/ab_product_20260921.jpg +3 -0
- assets/ab_spatial_20260920.jpg +3 -0
- assets/ab_spatial_20260921.jpg +3 -0
- assets/ab_typography_20260920.jpg +3 -0
- assets/ab_typography_20260921.jpg +3 -0
- assets/architecture_20260920_bf-bf.webp +3 -0
- assets/architecture_20260920_nv-nv.webp +3 -0
- assets/architecture_20260921_bf-bf.webp +3 -0
- assets/architecture_20260921_nv-nv.webp +3 -0
- assets/astronaut_20260920_bf-bf.webp +3 -0
- assets/astronaut_20260920_nv-nv.webp +3 -0
- assets/astronaut_20260921_bf-bf.webp +3 -0
- assets/astronaut_20260921_nv-nv.webp +3 -0
- assets/botanical_poster_20260920_bf-bf.webp +0 -0
- assets/botanical_poster_20260920_nv-nv.webp +0 -0
- assets/botanical_poster_20260921_bf-bf.webp +3 -0
- assets/botanical_poster_20260921_nv-nv.webp +0 -0
- assets/comparisons.json +170 -0
- assets/edit_portrait_bf-bf_extended_00001_.png +3 -0
- assets/edit_portrait_nv-nv_extended_00001_.png +3 -0
- assets/edit_product_bf-bf_extended_00001_.png +3 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,89 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
assets/01_Text_to_Image.png filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
assets/02_Image_Editing.png filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
assets/03_Transparent_RGBA.png filter=lfs diff=lfs merge=lfs -text
|
| 39 |
+
assets/04_2K_Typography.png filter=lfs diff=lfs merge=lfs -text
|
| 40 |
+
assets/ab_architecture_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
assets/ab_architecture_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 42 |
+
assets/ab_astronaut_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 43 |
+
assets/ab_astronaut_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 44 |
+
assets/ab_botanical_poster_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
assets/ab_botanical_poster_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 46 |
+
assets/ab_fashion_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 47 |
+
assets/ab_fashion_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 48 |
+
assets/ab_illustration_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 49 |
+
assets/ab_illustration_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 50 |
+
assets/ab_landscape_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 51 |
+
assets/ab_landscape_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 52 |
+
assets/ab_macro_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 53 |
+
assets/ab_macro_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 54 |
+
assets/ab_night_market_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 55 |
+
assets/ab_night_market_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 56 |
+
assets/ab_portrait_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 57 |
+
assets/ab_portrait_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 58 |
+
assets/ab_product_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 59 |
+
assets/ab_product_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 60 |
+
assets/ab_spatial_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 61 |
+
assets/ab_spatial_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 62 |
+
assets/ab_typography_20260920.jpg filter=lfs diff=lfs merge=lfs -text
|
| 63 |
+
assets/ab_typography_20260921.jpg filter=lfs diff=lfs merge=lfs -text
|
| 64 |
+
assets/architecture_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 65 |
+
assets/architecture_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 66 |
+
assets/architecture_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 67 |
+
assets/architecture_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 68 |
+
assets/astronaut_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 69 |
+
assets/astronaut_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 70 |
+
assets/astronaut_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 71 |
+
assets/astronaut_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 72 |
+
assets/botanical_poster_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 73 |
+
assets/edit_portrait_bf-bf_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 74 |
+
assets/edit_portrait_nv-nv_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 75 |
+
assets/edit_product_bf-bf_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 76 |
+
assets/edit_product_nv-nv_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 77 |
+
assets/editing.jpg filter=lfs diff=lfs merge=lfs -text
|
| 78 |
+
assets/fashion_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 79 |
+
assets/fashion_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 80 |
+
assets/fashion_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 81 |
+
assets/fashion_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 82 |
+
assets/heldout_chinese_bf-bf_heldout_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 83 |
+
assets/heldout_chinese_nv-nv_heldout_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 84 |
+
assets/heldout_two_references_bf-bf_heldout_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 85 |
+
assets/heldout_two_references_nv-nv_heldout_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 86 |
+
assets/heldout_wide_bf-bf_heldout_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 87 |
+
assets/heldout_wide_nv-nv_heldout_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 88 |
+
assets/illustration_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 89 |
+
assets/illustration_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 90 |
+
assets/illustration_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 91 |
+
assets/illustration_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 92 |
+
assets/landscape_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 93 |
+
assets/landscape_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 94 |
+
assets/landscape_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 95 |
+
assets/landscape_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 96 |
+
assets/macro_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 97 |
+
assets/macro_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 98 |
+
assets/macro_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 99 |
+
assets/macro_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 100 |
+
assets/night_market_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 101 |
+
assets/night_market_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 102 |
+
assets/night_market_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 103 |
+
assets/night_market_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 104 |
+
assets/portrait_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 105 |
+
assets/portrait_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 106 |
+
assets/portrait_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 107 |
+
assets/portrait_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 108 |
+
assets/product_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 109 |
+
assets/product_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 110 |
+
assets/product_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 111 |
+
assets/product_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 112 |
+
assets/transparency-checker.jpg filter=lfs diff=lfs merge=lfs -text
|
| 113 |
+
assets/transparent_fox_bf-bf_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 114 |
+
assets/transparent_fox_nv-nv_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 115 |
+
assets/typography_20260920_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 116 |
+
assets/typography_20260920_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 117 |
+
assets/typography_20260921_bf-bf.webp filter=lfs diff=lfs merge=lfs -text
|
| 118 |
+
assets/typography_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
|
| 119 |
+
assets/typography_2k_bf-bf_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 120 |
+
assets/typography_2k_nv-nv_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
|
| 121 |
+
input/prism_edit_reference.png filter=lfs diff=lfs merge=lfs -text
|
COMPARISONS.md
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# All matched comparisons
|
| 2 |
+
|
| 3 |
+
BF16 source weights versus the accelerated mixed NVFP4 stack. Same prompts and seeds, 1024x1024, 40 steps, Euler/simple, CFG 1.
|
| 4 |
+
|
| 5 |
+
| Subject | Seed | Side by side |
|
| 6 |
+
|---|---:|---|
|
| 7 |
+
| portrait | 20260920 | [Open](assets/ab_portrait_20260920.jpg) |
|
| 8 |
+
| portrait | 20260921 | [Open](assets/ab_portrait_20260921.jpg) |
|
| 9 |
+
| landscape | 20260920 | [Open](assets/ab_landscape_20260920.jpg) |
|
| 10 |
+
| landscape | 20260921 | [Open](assets/ab_landscape_20260921.jpg) |
|
| 11 |
+
| product | 20260920 | [Open](assets/ab_product_20260920.jpg) |
|
| 12 |
+
| product | 20260921 | [Open](assets/ab_product_20260921.jpg) |
|
| 13 |
+
| typography | 20260920 | [Open](assets/ab_typography_20260920.jpg) |
|
| 14 |
+
| typography | 20260921 | [Open](assets/ab_typography_20260921.jpg) |
|
| 15 |
+
| spatial | 20260920 | [Open](assets/ab_spatial_20260920.jpg) |
|
| 16 |
+
| spatial | 20260921 | [Open](assets/ab_spatial_20260921.jpg) |
|
| 17 |
+
| illustration | 20260920 | [Open](assets/ab_illustration_20260920.jpg) |
|
| 18 |
+
| illustration | 20260921 | [Open](assets/ab_illustration_20260921.jpg) |
|
| 19 |
+
| fashion | 20260920 | [Open](assets/ab_fashion_20260920.jpg) |
|
| 20 |
+
| fashion | 20260921 | [Open](assets/ab_fashion_20260921.jpg) |
|
| 21 |
+
| night market | 20260920 | [Open](assets/ab_night_market_20260920.jpg) |
|
| 22 |
+
| night market | 20260921 | [Open](assets/ab_night_market_20260921.jpg) |
|
| 23 |
+
| architecture | 20260920 | [Open](assets/ab_architecture_20260920.jpg) |
|
| 24 |
+
| architecture | 20260921 | [Open](assets/ab_architecture_20260921.jpg) |
|
| 25 |
+
| macro | 20260920 | [Open](assets/ab_macro_20260920.jpg) |
|
| 26 |
+
| macro | 20260921 | [Open](assets/ab_macro_20260921.jpg) |
|
| 27 |
+
| astronaut | 20260920 | [Open](assets/ab_astronaut_20260920.jpg) |
|
| 28 |
+
| astronaut | 20260921 | [Open](assets/ab_astronaut_20260921.jpg) |
|
| 29 |
+
| botanical poster | 20260920 | [Open](assets/ab_botanical_poster_20260920.jpg) |
|
| 30 |
+
| botanical poster | 20260921 | [Open](assets/ab_botanical_poster_20260921.jpg) |
|
| 31 |
+
|
| 32 |
+
Prompts and image mappings: [comparisons.json](assets/comparisons.json).
|
| 33 |
+
|
| 34 |
+
For the interactive slider, download gallery.html and assets together and open gallery.html locally.
|
PROMPTING.md
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Prompting Prism / Qwen Image 2.1
|
| 2 |
+
|
| 3 |
+
Start with a concrete natural-language paragraph: medium, subject, action,
|
| 4 |
+
relationships, composition, lighting, materials, then exact lettering. Keep each
|
| 5 |
+
attribute next to the object it describes. The supplied workflows use 40 steps,
|
| 6 |
+
Euler/simple, CFG 1, denoise 1, and fixed seeds.
|
| 7 |
+
|
| 8 |
+
## Text to image
|
| 9 |
+
|
| 10 |
+
Describe a scene that can be depicted in one image. For a product photo, specify
|
| 11 |
+
the object, material, label, supporting objects, surface, background and lighting.
|
| 12 |
+
For a portrait, specify the action and framing as well as the person's appearance.
|
| 13 |
+
The included product prompt is a complete starting point.
|
| 14 |
+
|
| 15 |
+
Set output width and height in Empty Latent Image. The encoder's resolution field
|
| 16 |
+
controls reference-image resizing, not the independent text-to-image canvas.
|
| 17 |
+
Changing image dimensions may change composition even with the same seed.
|
| 18 |
+
|
| 19 |
+
At CFG 1, the negative conditioning branch does not guide sampling. Put essential
|
| 20 |
+
requirements in the positive prompt. Higher CFG is not a validated improvement
|
| 21 |
+
for this package.
|
| 22 |
+
|
| 23 |
+
## Lettering
|
| 24 |
+
|
| 25 |
+
Quote the exact text and specify its position and hierarchy. Example:
|
| 26 |
+
|
| 27 |
+
> Large cream headline at the top reads exactly "TAKE THE SCENIC ROUTE".
|
| 28 |
+
> A smaller line at the bottom reads exactly "SLOW TRAVEL CLUB".
|
| 29 |
+
> Generous margins, flat navy, coral and cream inks, subtle paper grain.
|
| 30 |
+
|
| 31 |
+
Proofread the result. Clear headings and short labels worked in the release
|
| 32 |
+
examples, but the astronaut test also shows small-label spelling errors in the
|
| 33 |
+
BF16 baseline. Neither precision is a guarantee of perfect text.
|
| 34 |
+
|
| 35 |
+
## Image editing
|
| 36 |
+
|
| 37 |
+
Upload the reference in Load Image. Connect it and the matching VAE to
|
| 38 |
+
TextEncodeQwenImage21. Use that node's latent output for sampling, so the output
|
| 39 |
+
matches the processed first reference's size. The included editing workflow
|
| 40 |
+
already has those connections.
|
| 41 |
+
|
| 42 |
+
Name the changes and the elements to preserve:
|
| 43 |
+
|
| 44 |
+
> Change the amber glass bottle to deep translucent cobalt-blue glass and the
|
| 45 |
+
> brass cap to brushed silver. Replace the orange with lime and rosemary with
|
| 46 |
+
> lavender. Keep the bottle shape, cream EMBER label, marble pedestal, camera
|
| 47 |
+
> angle, background and lighting.
|
| 48 |
+
|
| 49 |
+
For multiple references, refer explicitly to the first and second images.
|
| 50 |
+
Reference preservation is semantic, not pixel-locked: small details, hand poses,
|
| 51 |
+
clothing coverage and object placement can change.
|
| 52 |
+
|
| 53 |
+
## Native transparency
|
| 54 |
+
|
| 55 |
+
Use an explicit RGBA instruction around the subject description:
|
| 56 |
+
|
| 57 |
+
> This is an RGBA image with transparency. A charming hand-painted gouache sticker
|
| 58 |
+
> of a small orange fox curled around a glowing blue book, round spectacles,
|
| 59 |
+
> moss-green scarf, clean complete silhouette and generous empty margin.
|
| 60 |
+
> The image has alpha channel and the background is transparent.
|
| 61 |
+
|
| 62 |
+
Save as PNG. Check the alpha channel or composite over a checkerboard. The hidden
|
| 63 |
+
RGB under transparent pixels can be colored; an RGB-only viewer is not an alpha
|
| 64 |
+
test. The supplied example has alpha values spanning 0 through 255.
|
| 65 |
+
|
| 66 |
+
## Comparisons
|
| 67 |
+
|
| 68 |
+
Keep the prompt, seed, dimensions, sampler, steps, backend and encoder mode fixed.
|
| 69 |
+
The optional NVFP4 encoder patch changes conditioning arithmetic and may change
|
| 70 |
+
the generated image. Compare stock and accelerated modes as distinct settings.
|
| 71 |
+
The gallery contains the complete 12-prompt, two-seed comparison set.
|
README.md
ADDED
|
@@ -0,0 +1,132 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: qwen-research-license
|
| 4 |
+
license_link: LICENSE
|
| 5 |
+
base_model: Comfy-Org/Qwen-Image-2.1
|
| 6 |
+
base_model_relation: quantized
|
| 7 |
+
pipeline_tag: text-to-image
|
| 8 |
+
tags:
|
| 9 |
+
- qwen-image
|
| 10 |
+
- qwen-image-2.1
|
| 11 |
+
- nvfp4
|
| 12 |
+
- comfyui
|
| 13 |
+
- diffusion-single-file
|
| 14 |
+
- quantized
|
| 15 |
+
- image-editing
|
| 16 |
+
- rtx-5090
|
| 17 |
+
---
|
| 18 |
+
|
| 19 |
+
# 🚀 Prism Image 2.1 — NVFP4 for ComfyUI
|
| 20 |
+
|
| 21 |
+
**The image model AND the text encoder. 12.42 GB for the complete package.**
|
| 22 |
+
|
| 23 |
+
I converted Qwen Image 2.1's image transformer and Qwen3-VL 8B language encoder
|
| 24 |
+
directly from the official BF16 weights into native ComfyUI NVFP4. The vision
|
| 25 |
+
tower, embeddings, output head, critical image-model tensors, and VAE stay BF16.
|
| 26 |
+
Real Blackwell FP4 kernels in both denoising and text conditioning, with an
|
| 27 |
+
included small runtime patch for the latter. No new nodes.
|
| 28 |
+
|
| 29 |
+
**Built with Qwen.** Research/evaluation use only under the included license.
|
| 30 |
+
|
| 31 |
+
**RTX 5090: 6.746 s median for a complete 1024×1024, 40-step run** versus
|
| 32 |
+
15.629 s with BF16 and 7.553 s with official INT8. Three warm runs per variant,
|
| 33 |
+
every node executed, including text encoding. This one-prompt benchmark used the
|
| 34 |
+
encoder patch; [full method and limits](VALIDATION.md).
|
| 35 |
+
|
| 36 |
+

|
| 37 |
+
|
| 38 |
+
## 📦 Download once, wire nothing
|
| 39 |
+
|
| 40 |
+
Tested with **ComfyUI 0.36.0**, PyTorch **2.14.0+cu130**, comfy-kitchen **0.2.35**,
|
| 41 |
+
and an **RTX 5090**. Use a ComfyUI build with `TextEncodeQwenImage21` and native
|
| 42 |
+
NVFP4 support. Older builds without the Qwen Image 2.1 integration will not work.
|
| 43 |
+
|
| 44 |
+
| File | Where it goes | Size |
|
| 45 |
+
|---|---|---:|
|
| 46 |
+
| [prism_image_2.1_nvfp4.safetensors](diffusion_models/prism_image_2.1_nvfp4.safetensors) | `ComfyUI/models/diffusion_models/` | 4.20 GB |
|
| 47 |
+
| [prism_qwen3vl_8b_nvfp4.safetensors](text_encoders/prism_qwen3vl_8b_nvfp4.safetensors) | `ComfyUI/models/text_encoders/` | 7.55 GB |
|
| 48 |
+
| [qwen_image_2.1_vae_bf16.safetensors](vae/qwen_image_2.1_vae_bf16.safetensors) | `ComfyUI/models/vae/` | 0.68 GB |
|
| 49 |
+
|
| 50 |
+
The original BF16 set is 32.44 GB. This package is **61.7% smaller on disk**.
|
| 51 |
+
File sizes are decimal GB, not a VRAM requirement. Keep your existing identical
|
| 52 |
+
VAE if you already have it.
|
| 53 |
+
|
| 54 |
+
1. Put the three files in the folders above and refresh ComfyUI's model list.
|
| 55 |
+
2. For accelerated NVFP4 text conditioning, apply the [runtime patch](runtime/README.md)
|
| 56 |
+
and restart. Without it, the same models and nodes work, but text conditioning
|
| 57 |
+
uses dequantized weights and FP32 activations. Denoising still uses NVFP4.
|
| 58 |
+
3. Import one of the workflows below using **Ctrl+O**, then click **Run**.
|
| 59 |
+
4. For editing, use **Upload** in Load Image to select the included
|
| 60 |
+
[reference image](input/prism_edit_reference.png), or your own image.
|
| 61 |
+
|
| 62 |
+
| Working workflow | Purpose |
|
| 63 |
+
|---|---|
|
| 64 |
+
| [01 — Text to image](workflows/01_Text_to_Image.json) | 1024×1024 product example; change the prompt and latent size |
|
| 65 |
+
| [02 — Image editing](workflows/02_Image_Editing.json) | One reference, coherent material/color/object changes |
|
| 66 |
+
| [03 — Transparent RGBA](workflows/03_Transparent_RGBA.json) | Native alpha output; no background-removal node |
|
| 67 |
+
| [04 — 2K typography](workflows/04_2K_Typography.json) | 2048×2048 poster with exact requested lettering |
|
| 68 |
+
|
| 69 |
+
The workflows use core `UNETLoader`, `CLIPLoader` (type **qwen_image**),
|
| 70 |
+
`TextEncodeQwenImage21`, `KSampler`, and VAE nodes. Leave diffusion weight dtype
|
| 71 |
+
at **default**. Defaults are 40 steps, Euler, simple scheduler, CFG 1, denoise 1.
|
| 72 |
+
The encoder node's resolution controls reference resizing; text-to-image output
|
| 73 |
+
size comes from Empty Latent Image. API equivalents are in [workflows/api](workflows/api).
|
| 74 |
+
|
| 75 |
+
## 👀 Same prompts, same seeds
|
| 76 |
+
|
| 77 |
+

|
| 78 |
+
|
| 79 |
+

|
| 80 |
+
|
| 81 |
+
All **24 matched pairs** are in [COMPARISONS.md](COMPARISONS.md). Download
|
| 82 |
+
`gallery.html` and its `assets` folder together for the offline comparison slider.
|
| 83 |
+
See [PROMPTING.md](PROMPTING.md) for text, editing, lettering and transparency tips.
|
| 84 |
+
|
| 85 |
+
See [VALIDATION.md](VALIDATION.md) for the test set, measured results, precision
|
| 86 |
+
tradeoffs, and execution requirements. The comparisons use BF16 source weights
|
| 87 |
+
against this complete accelerated NVFP4 package. Quantization changes results;
|
| 88 |
+
this is not a promise of pixel-identical output or a universal quality ranking.
|
| 89 |
+
|
| 90 |
+
## 🧠 What's actually quantized?
|
| 91 |
+
|
| 92 |
+
| Component | Policy |
|
| 93 |
+
|---|---|
|
| 94 |
+
| Image transformer | 192 attention Q/K/V/output and fused MLP gate/up/output matrices: NVFP4 |
|
| 95 |
+
| Qwen3-VL language backbone | 252 attention and MLP matrices: NVFP4 |
|
| 96 |
+
| Vision tower, token embeddings, language head | Original BF16 tensors retained |
|
| 97 |
+
| Image-model input/output, timestep, modulation, normalization | Original BF16 tensors retained |
|
| 98 |
+
| VAE | Original BF16 file, unchanged |
|
| 99 |
+
| Scales | E4M3 FP8 blocks of 16, FP32 tensor scales |
|
| 100 |
+
|
| 101 |
+
Native packed E2M1 FP4, native scale layout, and per-layer `comfy_quant` metadata.
|
| 102 |
+
This is a mixed-precision NVFP4 stack, not every parameter forced into four bits.
|
| 103 |
+
No GGUF loader, no requantization of the INT8 release, no dropped vision tower.
|
| 104 |
+
No calibration or fine-tuning was used for this conversion.
|
| 105 |
+
|
| 106 |
+
## 🔧 Reproduce it
|
| 107 |
+
|
| 108 |
+
Source repository: [Comfy-Org/Qwen-Image-2.1](https://huggingface.co/Comfy-Org/Qwen-Image-2.1),
|
| 109 |
+
revision `ace0edeb3791a594ddfa36ed5f41a178a394e921`.
|
| 110 |
+
|
| 111 |
+
Use the Python environment of the tested ComfyUI installation:
|
| 112 |
+
|
| 113 |
+
```powershell
|
| 114 |
+
python tools/convert.py path/to/qwen_image_2.1_bf16.safetensors diffusion_models/prism_image_2.1_nvfp4.safetensors --component dit --comfy-root path/to/ComfyUI
|
| 115 |
+
python tools/convert.py path/to/qwen3vl_8b_bf16.safetensors text_encoders/prism_qwen3vl_8b_nvfp4.safetensors --component encoder --comfy-root path/to/ComfyUI
|
| 116 |
+
```
|
| 117 |
+
|
| 118 |
+
The converter uses installed Comfy Kitchen operations, validates matrix alignment,
|
| 119 |
+
and emits tensor manifests with source/output hashes and reconstruction error.
|
| 120 |
+
`tools/verify.py SOURCE CONVERTED` checks the stored format, checkpoint hash, and
|
| 121 |
+
exact equality of every protected tensor. Final file checksums are in `SHA256SUMS`.
|
| 122 |
+
|
| 123 |
+
## ⚖️ License & credits
|
| 124 |
+
|
| 125 |
+
Qwen created the base model; Comfy-Org provided the native BF16 checkpoint package;
|
| 126 |
+
ComfyUI and Comfy Kitchen provide loading and GPU execution. Independent conversion,
|
| 127 |
+
workflow packaging, and RTX 5090 validation by **BennyDaBall_OG**.
|
| 128 |
+
|
| 129 |
+
The [Qwen Research License](LICENSE) limits use to non-commercial research and
|
| 130 |
+
evaluation; commercial use requires a separate license from Qwen. Read the full
|
| 131 |
+
license and [NOTICE](NOTICE). The included ComfyUI patch remains GPL-3.0.
|
| 132 |
+
This repo is not affiliated with Qwen or Comfy-Org.
|
SHA256SUMS
ADDED
|
@@ -0,0 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
dbfd3a6bf37d800ab2d5dfbb977fcc6e1b93f4c57ec41ff423598b93a6536ccf assets/01_Text_to_Image.png
|
| 2 |
+
4e6fa2ad6423a4bf308707d3d892c3616b614f63d634f4202816e92c8c048910 assets/02_Image_Editing.png
|
| 3 |
+
14dbba646932489c0bcbe513ec8151fa5d18005832202abe961d03544c8ef6f8 assets/03_Transparent_RGBA.png
|
| 4 |
+
95b2ecdf94cf414f51dbb4402176b3deaaad8d984797658ca013825770f6cdf1 assets/04_2K_Typography.png
|
| 5 |
+
22ae98495169b99a864775e7720630cb9efb3de9a99a1b436c209a8508d1eb9e assets/ab_architecture_20260920.jpg
|
| 6 |
+
5dc83f4493aabec649da97c4fccb43778e7d408e65f940b8a36b6f7253cde16f assets/ab_architecture_20260921.jpg
|
| 7 |
+
3657cd495de80e9a383f71b116aacfb3bb3b708f6cb5a84e6087dd9afd1bcb1d assets/ab_astronaut_20260920.jpg
|
| 8 |
+
8651d645ec839a90e97cef397492dd1c6628a0ea0827127a7c44c7ff8994bfcc assets/ab_astronaut_20260921.jpg
|
| 9 |
+
9c8b033d751b02ca57404323e0da59ddc130771246d38b5296491be6c2923d16 assets/ab_botanical_poster_20260920.jpg
|
| 10 |
+
f1249b09bed33b1955cc81ed068a165cfe0d7dc33d772e267e7c420085708e17 assets/ab_botanical_poster_20260921.jpg
|
| 11 |
+
47b64564650fbcc7311647f2cdf3b61fad50025babcc8ac22c2f0ffc02060791 assets/ab_fashion_20260920.jpg
|
| 12 |
+
7764788cb7287f1b4bd3f8572add42cb91518e3e0df96044c8536625575cda88 assets/ab_fashion_20260921.jpg
|
| 13 |
+
4ae372493b951d870c62c03eb1f082fa7ec6478846e7364d1154f851362f0f97 assets/ab_illustration_20260920.jpg
|
| 14 |
+
fcb1880d6e7affa6d0c51622c937e42fa1618354b6db510608a6295f8d8ef1c2 assets/ab_illustration_20260921.jpg
|
| 15 |
+
e55937e67bde8f14b2c7061cab4c18c738fb358b30fd9f04167148c9741de8a5 assets/ab_landscape_20260920.jpg
|
| 16 |
+
be6e963c3b087b915ff64a76be4897c7297fde82fe570e706136730709027157 assets/ab_landscape_20260921.jpg
|
| 17 |
+
940640b5ab0535405937dada3d0e88c55c83feb5cd4d4a8e6737b7ec2d022884 assets/ab_macro_20260920.jpg
|
| 18 |
+
b5b913b88b4b2d7ec8f37a7c08451a2a53ba6e965817eb7b49f87c46d3e18a0d assets/ab_macro_20260921.jpg
|
| 19 |
+
368181c2642d730adcc1b121041ee88ea5e25f23efe33d593982408ea50b97b5 assets/ab_night_market_20260920.jpg
|
| 20 |
+
5266ca07869dd84f6437a3f65ed3417c98404c5938f1341aeeb2d9766274d116 assets/ab_night_market_20260921.jpg
|
| 21 |
+
ccfe23331acbd05c5fcf8b0db46a2d1eeb1872e09a62b5800bac0d983f8f005f assets/ab_portrait_20260920.jpg
|
| 22 |
+
6c4804dac52ce29c2e31a84e00fcc5f7f2e11abcae595f1a7adef3d37a876bdc assets/ab_portrait_20260921.jpg
|
| 23 |
+
ab620447fc0e6b17e1bb47df61db047bd4ee15a9f7014c834af9abf595af7847 assets/ab_product_20260920.jpg
|
| 24 |
+
402220da9ae5f5d268fc9caa12f0decbd2d48164ad5453d031ce1b5dbc04b318 assets/ab_product_20260921.jpg
|
| 25 |
+
7498b7a7a3e766c04f27d5d28302f69c10f5c75bc7e7b562172bcd169efd03e3 assets/ab_spatial_20260920.jpg
|
| 26 |
+
b3faee6e319a5c675c91167ef0704b63580e9dc8d0340550afeb0c72759b6792 assets/ab_spatial_20260921.jpg
|
| 27 |
+
adeb0e4ed51ce0ff6a53793e3c301f4d29713f23c30ba21c4c0f4a2b849bb0d3 assets/ab_typography_20260920.jpg
|
| 28 |
+
b1be26c4fa502e0a1fba47227573b932b4fe38d5aa8c61234392b40c94488949 assets/ab_typography_20260921.jpg
|
| 29 |
+
81da8fc0e71b2b3373c1cf46734ca449d5f0ba049c77921f0c28e65d19b58a80 assets/architecture_20260920_bf-bf.webp
|
| 30 |
+
bd505f621203fa42fc3c0521d1977ec3d2b7899db92c5bad1c8148c9c40a8f2a assets/architecture_20260920_nv-nv.webp
|
| 31 |
+
468e043ba915698ccd77c67f5fae8ed53c0cd9219aed2f2465b7d658a8e6c3e8 assets/architecture_20260921_bf-bf.webp
|
| 32 |
+
f1217e758568946787a198d3fb5f7f3e6284bb662f68549108962862248e242a assets/architecture_20260921_nv-nv.webp
|
| 33 |
+
578215e7ebc5d59040d7b12725fc23417d81e1f16c3b96ef9170cca1207e7884 assets/astronaut_20260920_bf-bf.webp
|
| 34 |
+
fd893ec06228f619a706695664b76ff5d66d08d937e5f9b851f1faec30ea0793 assets/astronaut_20260920_nv-nv.webp
|
| 35 |
+
4beb4c3ec852bd711d5cc5bddb7b1e8f04f35f98025be92117f27c6a197b8807 assets/astronaut_20260921_bf-bf.webp
|
| 36 |
+
819054921f85945f531328fe5347abc76c9bd92d97e0d312878f0418c4b49b91 assets/astronaut_20260921_nv-nv.webp
|
| 37 |
+
5a4cc9c7cc07ca44e24d80d2522dbdf0b96b82d8d0e1895a8feaaad965bcaaa7 assets/botanical_poster_20260920_bf-bf.webp
|
| 38 |
+
80e35caa110283ce376014e7b0a500a753c322f7f32be86c179c2b61c1d95e7b assets/botanical_poster_20260920_nv-nv.webp
|
| 39 |
+
b756961e69efadf8f66cf13784092777bf00429bbc194606562f4b4850aba697 assets/botanical_poster_20260921_bf-bf.webp
|
| 40 |
+
7c180e23340bab3f4a3e7c0f4293769fa1d23c532b6aac157cf2f2db82298c43 assets/botanical_poster_20260921_nv-nv.webp
|
| 41 |
+
5567516bc88d9e486a86b44e309f1e80d36bd09facde38c46f8b64a5644d8320 assets/comparisons.json
|
| 42 |
+
883d3436df3c78142c7590f4b1c73518308498a98bb3b1343cf0ef2c80b85ad2 assets/edit_portrait_bf-bf_extended_00001_.png
|
| 43 |
+
b6a97056d829de05da472a06c8c1a7e199a4866ff0db1fed2597431f281a624c assets/edit_portrait_nv-nv_extended_00001_.png
|
| 44 |
+
a0117e505673ca323e1c358ab5b6532ccc1fdcf9e08c87ff3a43712f50f0b501 assets/edit_product_bf-bf_extended_00001_.png
|
| 45 |
+
4e6fa2ad6423a4bf308707d3d892c3616b614f63d634f4202816e92c8c048910 assets/edit_product_nv-nv_extended_00001_.png
|
| 46 |
+
f05044570f14a5c4b75ef5803040b6cf5e8104640da248b64ffecf64c730b2f9 assets/editing.jpg
|
| 47 |
+
ef6482fc15134a4ef537883911328b5e209a63d04da2a266972ebf1d435df0e5 assets/fashion_20260920_bf-bf.webp
|
| 48 |
+
d7c370c533559d7752cae2593602aa2e6ef615b2ef3f0132686b8cfae6bf034a assets/fashion_20260920_nv-nv.webp
|
| 49 |
+
82418e7dcbb5f8c7e48cd74715ce4f5acfa646d51e3de33bb177b945ad377926 assets/fashion_20260921_bf-bf.webp
|
| 50 |
+
ae268aebdd545da33d5385c8578d8eae59ea0717e1cd099847beb1c6d07b8f36 assets/fashion_20260921_nv-nv.webp
|
| 51 |
+
31d35c1c210cddeb0ab3d2e44b99fb29d5bf0c45c3ce7ed7c118a2f6c602815f assets/heldout_chinese_bf-bf_heldout_00001_.png
|
| 52 |
+
26e172341a68f36f2465e6eb981795a748e7dceddf62cea99425fe8e1583910c assets/heldout_chinese_nv-nv_heldout_00001_.png
|
| 53 |
+
d6adcace2ef4b50a1209130d1808d5b295e9cba0ba685c3b5ef72b4c387e9a6d assets/heldout_two_references_bf-bf_heldout_00001_.png
|
| 54 |
+
e4537d06aa174e9be55087de06dde787c87872fceaf0c8db19833d5304eb35d2 assets/heldout_two_references_nv-nv_heldout_00001_.png
|
| 55 |
+
c8920393163031b29666b8f26612ff2b881cb198bee2d457bfe3838927b9d214 assets/heldout_wide_bf-bf_heldout_00001_.png
|
| 56 |
+
bab5d67a50bdd4077bb3338ca3f14c6478c062359e3003b64b6e05fc5b61574b assets/heldout_wide_nv-nv_heldout_00001_.png
|
| 57 |
+
a5e332a824c8b7612a09cbdacbf5da2f20e98da8bdafbad4d45809c1eb431d7b assets/illustration_20260920_bf-bf.webp
|
| 58 |
+
3559d664dcc6c9c6bf23da67629565fbb249d9b11c141b1c7dfc228dbb8d56be assets/illustration_20260920_nv-nv.webp
|
| 59 |
+
a5404ab62c762f40524d5f83346fc984a2173f13c507c822066fd0b30e3fb53d assets/illustration_20260921_bf-bf.webp
|
| 60 |
+
72006e889343dc7f97f36263aae353e47bd754798f029bff4cea01dba4731355 assets/illustration_20260921_nv-nv.webp
|
| 61 |
+
1223892f81fa4c2964fcf2072a90f701c571ecd49c25cdf10a94930d87ed41b5 assets/landscape_20260920_bf-bf.webp
|
| 62 |
+
a055f63333de2b53afc533756a0e47fe33a13807ce4e55913b2c5668a95aa21d assets/landscape_20260920_nv-nv.webp
|
| 63 |
+
8549bc53a0c8ab1c5989165819af8f988fe21fa97a527e4343a2f15565961986 assets/landscape_20260921_bf-bf.webp
|
| 64 |
+
f55646dd998793b332f9e93ba5bd46a318c7e830d14243011538962594a84cf7 assets/landscape_20260921_nv-nv.webp
|
| 65 |
+
a168fd08b430ee107d8dee6f19c12d8b957b09387a40eebeeb5ef840fa9a4443 assets/macro_20260920_bf-bf.webp
|
| 66 |
+
a7f7f26dd3e9449034b9c900392a3489cd7e1c57f61d770959a0c7066e2257d0 assets/macro_20260920_nv-nv.webp
|
| 67 |
+
9b634ba6122e645091f33bee0b8ee5dc89e57911e0297899f55ebe59c6228701 assets/macro_20260921_bf-bf.webp
|
| 68 |
+
71ddd83850a07a8d22d5b5429260fbdfc8410bf8210cbf0e31cfa4b7caed651e assets/macro_20260921_nv-nv.webp
|
| 69 |
+
5ee877a57b21ffa44f9ac23795ee0cbab6508fcd1779f772a586c17822795434 assets/night_market_20260920_bf-bf.webp
|
| 70 |
+
73e74351d162928b36131cffd6d3835e60940af43f1610567107352aa33bdcff assets/night_market_20260920_nv-nv.webp
|
| 71 |
+
a5aa39a5e0990c9bf8fbc46eac73e5cec1b99d2b380b786e62f8ec19886a7ce0 assets/night_market_20260921_bf-bf.webp
|
| 72 |
+
4410d1943de8641410309612ada3a951de7a59b25576e5fa7d72148ccafba0f9 assets/night_market_20260921_nv-nv.webp
|
| 73 |
+
1151d175b1be1a3aef777c575959c378b643bac3da6ee712864eb805678cad7e assets/portrait_20260920_bf-bf.webp
|
| 74 |
+
c2d8022b1f9cb578c42756ed88540249fb1d330fc93d3a9d8045414e0ff6323f assets/portrait_20260920_nv-nv.webp
|
| 75 |
+
80fc434a36fdc66fd27e042ae6c599d841786e473d04476dee6724681b04be48 assets/portrait_20260921_bf-bf.webp
|
| 76 |
+
856fada5c8c947c14db2b296682292a48c51f21dd422ece2cb3e37c4c28990f9 assets/portrait_20260921_nv-nv.webp
|
| 77 |
+
da24b7dc383ab6d6f718fee60a7622500ffc2d35fad80707216ba50d8cbb8963 assets/product_20260920_bf-bf.webp
|
| 78 |
+
24db6ddea7ff8f443a3d97967c052fb4cc4c7287215e650db32b9e6e22d8007a assets/product_20260920_nv-nv.webp
|
| 79 |
+
bd42934f8fbc3f1a29e1aaa4f44fac23042e438d6cbfca8341c5c8050df607b9 assets/product_20260921_bf-bf.webp
|
| 80 |
+
739ba08a8b710b548081f2229f81371b704fea4163e17ae000899fdd6afd041b assets/product_20260921_nv-nv.webp
|
| 81 |
+
240498636eb56531931b1e51771de7a86483837e8b8455e9a050cc5bfebcd497 assets/spatial_20260920_bf-bf.webp
|
| 82 |
+
5fc39e36d82ac1f37c2461ea29cf9dfa7af25e4f22aab89e433042ed44edad27 assets/spatial_20260920_nv-nv.webp
|
| 83 |
+
4b17c4fac64b3053c3dfc5762081e54b764b1765e7c2c9144a550aaf27e31c7b assets/spatial_20260921_bf-bf.webp
|
| 84 |
+
e2a09ca565ca07e3c433a38ae01af090a743709fb73ee85071ffd34ff6b193c7 assets/spatial_20260921_nv-nv.webp
|
| 85 |
+
637d0128f8d1b3ba861601bea2321783526d55318786553fae5d52148862f5dc assets/transparency-checker.jpg
|
| 86 |
+
d05dfcb96e822c4dc7659cc618b87ccb4821111d99f2743a8531afc56a79d545 assets/transparent_fox_bf-bf_extended_00001_.png
|
| 87 |
+
14dbba646932489c0bcbe513ec8151fa5d18005832202abe961d03544c8ef6f8 assets/transparent_fox_nv-nv_extended_00001_.png
|
| 88 |
+
ef6723ce3d0814ed25a54ff13fb1bc3efc38c7384d78af66122d78835cc665c0 assets/typography_20260920_bf-bf.webp
|
| 89 |
+
45019509e84a9388b2357aa9dbbdab5b94dc23002dcad8f37bd1ebe5d8497670 assets/typography_20260920_nv-nv.webp
|
| 90 |
+
92bce0f6454257178061c66d520c34339a480b25f36128b4034b03ff3796d4d4 assets/typography_20260921_bf-bf.webp
|
| 91 |
+
14742b11cf638a2c86fc476b389711e5597be36e660679fa1d89c6179caa1be2 assets/typography_20260921_nv-nv.webp
|
| 92 |
+
6eeab96fd9b89a5c83cc945e7799c9a16b37b503cd4b961dd2364e339a7cacf0 assets/typography_2k_bf-bf_extended_00001_.png
|
| 93 |
+
95b2ecdf94cf414f51dbb4402176b3deaaad8d984797658ca013825770f6cdf1 assets/typography_2k_nv-nv_extended_00001_.png
|
| 94 |
+
f1352b74e5fa07594113b4fe69a914e8ad36e8dd7ce16a27b1f46df0d3098cfe COMPARISONS.md
|
| 95 |
+
50bfc0c644df3f661447024c380c9c433c8f0f5edcc4878b6695dcab533c7f1b diffusion_models/prism_image_2.1_nvfp4.manifest.json
|
| 96 |
+
4fdbaec94c19f7b8888f1e1ddbae05354c014492f9b2547b1cd8dca0ce923526 diffusion_models/prism_image_2.1_nvfp4.safetensors
|
| 97 |
+
a0f54997ba63caa3540823bd5e7e6356f8c3ea16cc796b32477eb45b000066bf gallery.html
|
| 98 |
+
936375ea3fb35c7c33dbe59c7090c35edf019e6119bc6b775f74f3d859b49a54 input/prism_edit_reference.png
|
| 99 |
+
06b7d356cd8cb6b79aa2eef3f934d1e1e8e1c2aa1e85dcc23915dd9a68197249 LICENSE
|
| 100 |
+
c303de7acfe67db47fceaa8599c156895e462c06482dcd6b1e61d46ce02cadf3 NOTICE
|
| 101 |
+
8a8fafb7836a4d8eb02d176a3beef5d4dd522eff1084661c941494f5feba0f23 PROMPTING.md
|
| 102 |
+
6b8622eb6d6adfd66d0545c327019519ac1a491e878be4d99da46cad5aece479 README.md
|
| 103 |
+
230184f60bae2feaf244f10a8bac053c8ff33a183bcc365b4d8b876d2b7f4809 runtime/LICENSE-GPL-3.0
|
| 104 |
+
2c13ec9046ffa720484c569acce153f4759da47e9bc008af56d68d583854e2fa runtime/qwen21-nvfp4-conditioning.patch
|
| 105 |
+
524bf47b1f5b061514314d16f6afa6a02d21232545e3e96297c00781fc67733a runtime/README.md
|
| 106 |
+
41ac84507beed8d4bf5f46f7c49fa91c12162f73a354598b9943888411d57381 text_encoders/prism_qwen3vl_8b_nvfp4.manifest.json
|
| 107 |
+
cdd9b85bcb60d5cb358ddecfd45a48fa15cb728b70d88bd2028840eaef095d16 text_encoders/prism_qwen3vl_8b_nvfp4.safetensors
|
| 108 |
+
c9b72630ba5c3a6bb842afe948a778ae7d742e92d94407310976e8c0f2f7f27d tools/convert.py
|
| 109 |
+
b5648e901612c6077ed55b93192bf3a152f842736e08b859082af6635e2339f9 tools/run_workflow.py
|
| 110 |
+
4363e4224c409a8c8a3099a06e75d9881206fbe4f541d44bde05dc872de2c3dc tools/verify.py
|
| 111 |
+
bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9 vae/qwen_image_2.1_vae_bf16.safetensors
|
| 112 |
+
2a6b6c514cce8041f313e7a68d9e344560b1ebf68feefb576a0c1bb627db0a94 validation/alpha.json
|
| 113 |
+
24f0362777e6970f4d71085968478deb5b34e4116c678b11573ad637a66ac3b2 validation/benchmark.json
|
| 114 |
+
135690b380fbe6e4f0362dd5434abcb5736763848a224417bdc556228f39f235 validation/cases.json
|
| 115 |
+
a8b8f8ad67db63da99a7dc20701dfdb74235c847d9985c9cd364103f8f6de709 validation/frontend-workflows.json
|
| 116 |
+
6892ea5bd1fb4745f677c7e38f6e117b6daa813ad694500b3fd79e32e654e95f validation/kernel_audit.jsonl
|
| 117 |
+
da5ab176ad20f4cdfb3994511eea8c8ce8ceb025d4d97c3e3dee6d2041670feb validation/roundtrip.json
|
| 118 |
+
1fd3f3e9d53fbef667628d34057b056d2f6c7f62fdbfe2230bbac1ba77635f38 VALIDATION.md
|
| 119 |
+
9a4ef8c30464cc453a29914904e01c5096051a35ee50cfc63fb9e0d968a07c00 workflows/01_Text_to_Image.json
|
| 120 |
+
42b78026114db814333ecbd7c86117f32f61a736fe630dad67950d9bfd3fec28 workflows/02_Image_Editing.json
|
| 121 |
+
df44abe670d355ff885477364f206d70c4086c13df85a1705129ad6247f05524 workflows/03_Transparent_RGBA.json
|
| 122 |
+
be45564d77c151fa92fbd3420cc7b62eb428d8e9d65d1755fe72057e1a9c16f7 workflows/04_2K_Typography.json
|
| 123 |
+
13a543282fae430f5a37b70688741300d022eb9a7adea02396898055c5d273de workflows/api/01_Text_to_Image.api.json
|
| 124 |
+
68224d0a803e823f78cf82dc636a6881b75671348b8cdd300f7cf8ac7493c241 workflows/api/02_Image_Editing.api.json
|
| 125 |
+
cc742cfda9d55f9ecd5dbd2a6aeb87a8726d8c152f66c0f4b54b629a9ad4e3e2 workflows/api/03_Transparent_RGBA.api.json
|
| 126 |
+
ab18f014220866315d52dbb5ff7e5a6f42bd95d6b70219a8c828266b468efff5 workflows/api/04_2K_Typography.api.json
|
VALIDATION.md
ADDED
|
@@ -0,0 +1,125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Validation — RTX 5090, 20 September 2026
|
| 2 |
+
|
| 3 |
+
## What passed
|
| 4 |
+
|
| 5 |
+
- Both native safetensors checkpoints loaded through standard ComfyUI loaders.
|
| 6 |
+
- 48 main renders: 12 prompts × two seeds × BF16/NVFP4, all at 1024×1024.
|
| 7 |
+
- Eight extended renders: product edit, portrait edit, transparent RGBA and 2K
|
| 8 |
+
typography, each with BF16 and NVFP4.
|
| 9 |
+
- Six held-out renders: Chinese lettering, 1536×864 landscape and two-reference
|
| 10 |
+
image editing, each with BF16 and NVFP4. These prompts were not used to choose
|
| 11 |
+
the conversion recipe.
|
| 12 |
+
- Four shipped workflows imported and executed through the actual frontend.
|
| 13 |
+
- Three subjects tested across all four combinations of BF16/NVFP4 transformer
|
| 14 |
+
and encoder weights, using the original stock conditioning path.
|
| 15 |
+
- Real SM120 block-scaled FP4 CUDA kernels observed in both language conditioning
|
| 16 |
+
and image denoising with the accelerated encoder patch.
|
| 17 |
+
- All 571 protected tensors checked for exact equality against the source files.
|
| 18 |
+
All 444 quantized matrices checked for packed shape, marker, scale dtype and
|
| 19 |
+
finite scales. Source and release file hashes recorded.
|
| 20 |
+
- Hub-uploaded weights downloaded into a separate directory and matched against
|
| 21 |
+
the release SHA-256 hashes. The VAE hash matches the official original.
|
| 22 |
+
- All four API workflows also passed using the Hub-downloaded files on the
|
| 23 |
+
unpatched base runtime, with custom nodes disabled. Thus stock compatibility
|
| 24 |
+
and optional accelerated conditioning were tested separately.
|
| 25 |
+
|
| 26 |
+
No calibration, optimization against the test prompts, or fine-tuning was used.
|
| 27 |
+
The quantization recipe remained unchanged after its first conversion.
|
| 28 |
+
|
| 29 |
+
## Measured performance
|
| 30 |
+
|
| 31 |
+
1024×1024 EMBER product prompt from the shipped generation workflow, seed
|
| 32 |
+
20260922, 40 steps, Euler/simple, CFG 1, denoise 1. One first-after-unload run,
|
| 33 |
+
then three warm runs per variant. **Every node executed on every repeat** using
|
| 34 |
+
`--cache-none`; the histories were checked for an empty cached-node list.
|
| 35 |
+
Elapsed time comes from server execution-start/success timestamps, not polling
|
| 36 |
+
latency. Includes conditioning, denoising, VAE decode and PNG save.
|
| 37 |
+
|
| 38 |
+
| Weights / encoder mode | Warm median | Warm range | First after unload | Largest sampled GPU use, warm |
|
| 39 |
+
|---|---:|---:|---:|---:|
|
| 40 |
+
| Official BF16 / stock FP32 conditioning | 15.629 s | 15.553–15.650 s | 18.932 s | 31,677 MiB |
|
| 41 |
+
| Official INT8 ConvRot / stock conditioning | 7.553 s | 7.496–7.591 s | 9.973 s | 18,433 MiB |
|
| 42 |
+
| Prism NVFP4 / patched BF16 + NVFP4 conditioning | **6.746 s** | **6.706–6.749 s** | **8.083 s** | **13,249 MiB** |
|
| 43 |
+
|
| 44 |
+
This NVFP4 workflow delivered **2.32× the BF16 throughput** and took **10.7% less
|
| 45 |
+
time than INT8** in this one controlled test. These are local measurements, not
|
| 46 |
+
general speed guarantees. First-after-unload is not a cold filesystem-cache or
|
| 47 |
+
fresh-boot test. The quality-suite runs used ordinary node caching and are not
|
| 48 |
+
the basis of this performance table.
|
| 49 |
+
|
| 50 |
+
GPU memory was sampled through `nvidia-smi` every 100 ms. Values are **whole-GPU
|
| 51 |
+
usage**, including Windows and other idle desktop services, not PyTorch peak
|
| 52 |
+
allocated memory or a minimum VRAM requirement. The largest sampled NVFP4 warm
|
| 53 |
+
value is about 12.94 GiB. No other model inference job was running during this
|
| 54 |
+
benchmark. Hardware other than the RTX 5090 was not benchmarked.
|
| 55 |
+
|
| 56 |
+
Raw measurements: [benchmark.json](validation/benchmark.json).
|
| 57 |
+
|
| 58 |
+
## Quality observations and limits
|
| 59 |
+
|
| 60 |
+
The main set covers portraits, landscapes, glass products, typography, object
|
| 61 |
+
counts and spatial relationships, illustration, fashion, night scenes,
|
| 62 |
+
architecture, macro, astronaut portraits and botanical posters. All 24 matched
|
| 63 |
+
pairs are published in [COMPARISONS.md](COMPARISONS.md).
|
| 64 |
+
|
| 65 |
+
Visual review found coherent images and preserved useful text, material and
|
| 66 |
+
editing capabilities. It also found real differences: changed portrait details,
|
| 67 |
+
object placement, typography layout and small material/color details. In one
|
| 68 |
+
macro pair the quantized spider is less green. Small astronaut-label spelling
|
| 69 |
+
was imperfect in the BF16 baseline too. The Chinese probe retained the requested
|
| 70 |
+
wording, but NVFP4 moved the subtitle above the cup rather than below it.
|
| 71 |
+
|
| 72 |
+
The two-reference test combined the reference woman with the EMBER bottle and
|
| 73 |
+
removed the violin, while changing hand pose. The product edit preserved the
|
| 74 |
+
label and scene while changing glass, metal and props. Portrait recoloring can
|
| 75 |
+
also change sleeve coverage. This is semantic reference preservation, not a
|
| 76 |
+
pixel-locked edit.
|
| 77 |
+
|
| 78 |
+
The transparent example has genuine RGBA output: alpha minimum 0, maximum 255,
|
| 79 |
+
49.3% of pixels below alpha 16 and 49.7% above alpha 240. See the original
|
| 80 |
+
[RGBA PNG](assets/03_Transparent_RGBA.png) and the
|
| 81 |
+
[checkerboard composite](assets/transparency-checker.jpg). RGB-only viewers may
|
| 82 |
+
show the purple RGB values stored underneath transparent pixels.
|
| 83 |
+
|
| 84 |
+
This is a small local evaluation, not a blinded preference study or a benchmark
|
| 85 |
+
proving no quality loss. There was no LoRA, ControlNet, training or video test.
|
| 86 |
+
Long text, exhaustive multilingual coverage and every reference-image count were
|
| 87 |
+
not tested. Stock encoder arithmetic and accelerated encoder arithmetic can
|
| 88 |
+
produce different images even with identical weights and seeds.
|
| 89 |
+
|
| 90 |
+
## Runtime and kernel evidence
|
| 91 |
+
|
| 92 |
+
- Windows, Python 3.13.12, RTX 5090 32 GB.
|
| 93 |
+
- ComfyUI 0.36.0, base commit `99073836d45f66053c45ba8564984e6def9cebba`.
|
| 94 |
+
- Included Qwen-specific encoder patch; local validation commit `2d324220`.
|
| 95 |
+
- PyTorch `2.14.0+cu130`; comfy-kitchen `0.2.35`; comfy-aimdo `0.5.5`.
|
| 96 |
+
- Frontend package `1.53.6`; safetensors `0.8.0`.
|
| 97 |
+
- Comfy Kitchen attention, normal VRAM mode, async offload, batch size one.
|
| 98 |
+
- All tests ran with `--disable-all-custom-nodes`.
|
| 99 |
+
|
| 100 |
+
The profiler captured kernels beginning
|
| 101 |
+
`cutlass3x_sm120_bstensorop_s16864gemm_block_scaled_ue4m3xe2m1_ue4m3xe2m1`
|
| 102 |
+
for short conditioning sequences and image-denoising sequences. This confirms
|
| 103 |
+
hardware FP4 execution, rather than relying on checkpoint names or successful
|
| 104 |
+
loading. Raw observations: [kernel_audit.jsonl](validation/kernel_audit.jsonl).
|
| 105 |
+
No NVFP4 matmul-fallback warning occurred in the successful audited runs.
|
| 106 |
+
|
| 107 |
+
The first acceleration probe exposed FP32 activations entering a BF16/FP16-only
|
| 108 |
+
quantizer. The included patch fixes that at the Qwen Image 2.1 encoding boundary:
|
| 109 |
+
vision preprocessing stays unchanged, the resulting multimodal embeddings become
|
| 110 |
+
BF16 only for NVFP4 conditioning on supported hardware, and existing ComfyUI
|
| 111 |
+
matmul dispatch is used. BF16 and INT8 checkpoints retain their original path.
|
| 112 |
+
|
| 113 |
+
## Precision and provenance
|
| 114 |
+
|
| 115 |
+
The DiT converts 192 matrices and retains its other 73 tensors. The encoder
|
| 116 |
+
converts 252 matrices and retains its other 498 tensors, including the complete
|
| 117 |
+
vision tower, embeddings and output head. Per-matrix relative reconstruction
|
| 118 |
+
RMSE is recorded in the manifests; it is not an image-quality score.
|
| 119 |
+
|
| 120 |
+
Source revision: `Comfy-Org/Qwen-Image-2.1@ace0edeb3791a594ddfa36ed5f41a178a394e921`.
|
| 121 |
+
The image transformer source SHA-256 is
|
| 122 |
+
`89f4158d066cc33906a199fca85634f766892dd78f49b6698dabf187ac86c4bc`.
|
| 123 |
+
The encoder source SHA-256 is
|
| 124 |
+
`68bdc82bc1b66851162ae656225e7e2068166b603db19bd5d5a3b90eb12669a9`.
|
| 125 |
+
Full tensor manifests sit beside each checkpoint; `SHA256SUMS` covers the release.
|
assets/01_Text_to_Image.png
ADDED
|
Git LFS Details
|
assets/02_Image_Editing.png
ADDED
|
Git LFS Details
|
assets/03_Transparent_RGBA.png
ADDED
|
Git LFS Details
|
assets/04_2K_Typography.png
ADDED
|
Git LFS Details
|
assets/ab_architecture_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_architecture_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_astronaut_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_astronaut_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_botanical_poster_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_botanical_poster_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_fashion_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_fashion_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_illustration_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_illustration_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_landscape_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_landscape_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_macro_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_macro_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_night_market_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_night_market_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_portrait_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_portrait_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_product_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_product_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_spatial_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_spatial_20260921.jpg
ADDED
|
Git LFS Details
|
assets/ab_typography_20260920.jpg
ADDED
|
Git LFS Details
|
assets/ab_typography_20260921.jpg
ADDED
|
Git LFS Details
|
assets/architecture_20260920_bf-bf.webp
ADDED
|
Git LFS Details
|
assets/architecture_20260920_nv-nv.webp
ADDED
|
Git LFS Details
|
assets/architecture_20260921_bf-bf.webp
ADDED
|
Git LFS Details
|
assets/architecture_20260921_nv-nv.webp
ADDED
|
Git LFS Details
|
assets/astronaut_20260920_bf-bf.webp
ADDED
|
Git LFS Details
|
assets/astronaut_20260920_nv-nv.webp
ADDED
|
Git LFS Details
|
assets/astronaut_20260921_bf-bf.webp
ADDED
|
Git LFS Details
|
assets/astronaut_20260921_nv-nv.webp
ADDED
|
Git LFS Details
|
assets/botanical_poster_20260920_bf-bf.webp
ADDED
|
assets/botanical_poster_20260920_nv-nv.webp
ADDED
|
assets/botanical_poster_20260921_bf-bf.webp
ADDED
|
Git LFS Details
|
assets/botanical_poster_20260921_nv-nv.webp
ADDED
|
assets/comparisons.json
ADDED
|
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"name": "portrait",
|
| 4 |
+
"seed": 20260920,
|
| 5 |
+
"prompt": "Editorial environmental photograph of an elderly female violin maker in her sunlit wooden workshop. She holds a small carving tool in her right hand and steadies an unfinished violin with her left hand. Natural skin texture, thoughtful expression, curled wood shavings on the workbench, soft window light from the left, warm subdued colors, shallow depth of field, 50mm lens. Waist-up composition, candid and believable.",
|
| 6 |
+
"bf": "portrait_20260920_bf-bf.webp",
|
| 7 |
+
"nv": "portrait_20260920_nv-nv.webp"
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"name": "portrait",
|
| 11 |
+
"seed": 20260921,
|
| 12 |
+
"prompt": "Editorial environmental photograph of an elderly female violin maker in her sunlit wooden workshop. She holds a small carving tool in her right hand and steadies an unfinished violin with her left hand. Natural skin texture, thoughtful expression, curled wood shavings on the workbench, soft window light from the left, warm subdued colors, shallow depth of field, 50mm lens. Waist-up composition, candid and believable.",
|
| 13 |
+
"bf": "portrait_20260921_bf-bf.webp",
|
| 14 |
+
"nv": "portrait_20260921_nv-nv.webp"
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"name": "landscape",
|
| 18 |
+
"seed": 20260920,
|
| 19 |
+
"prompt": "Cinematic landscape photograph of a tiny red mountain refuge on a rocky ridge above a sea of clouds at sunrise. A winding stone footpath leads from the bottom left toward the refuge. Jagged distant peaks, warm sunlight grazing the roof, cool blue shadows, thin luminous mist, realistic geology, restrained natural colors. Wide composition with a strong sense of scale, no people, no lettering.",
|
| 20 |
+
"bf": "landscape_20260920_bf-bf.webp",
|
| 21 |
+
"nv": "landscape_20260920_nv-nv.webp"
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"name": "landscape",
|
| 25 |
+
"seed": 20260921,
|
| 26 |
+
"prompt": "Cinematic landscape photograph of a tiny red mountain refuge on a rocky ridge above a sea of clouds at sunrise. A winding stone footpath leads from the bottom left toward the refuge. Jagged distant peaks, warm sunlight grazing the roof, cool blue shadows, thin luminous mist, realistic geology, restrained natural colors. Wide composition with a strong sense of scale, no people, no lettering.",
|
| 27 |
+
"bf": "landscape_20260921_bf-bf.webp",
|
| 28 |
+
"nv": "landscape_20260921_nv-nv.webp"
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"name": "product",
|
| 32 |
+
"seed": 20260920,
|
| 33 |
+
"prompt": "Premium studio product photograph of one amber glass perfume bottle with a brushed brass cap, standing on a dark green marble block. A small cream paper label on the bottle reads \"EMBER\" in elegant black capital letters. A sliced orange and a sprig of rosemary rest beside the block. Warm rim lighting, soft front fill, rich realistic materials, controlled glass reflections, clean charcoal background, centered luxury advertising composition.",
|
| 34 |
+
"bf": "product_20260920_bf-bf.webp",
|
| 35 |
+
"nv": "product_20260920_nv-nv.webp"
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"name": "product",
|
| 39 |
+
"seed": 20260921,
|
| 40 |
+
"prompt": "Premium studio product photograph of one amber glass perfume bottle with a brushed brass cap, standing on a dark green marble block. A small cream paper label on the bottle reads \"EMBER\" in elegant black capital letters. A sliced orange and a sprig of rosemary rest beside the block. Warm rim lighting, soft front fill, rich realistic materials, controlled glass reflections, clean charcoal background, centered luxury advertising composition.",
|
| 41 |
+
"bf": "product_20260921_bf-bf.webp",
|
| 42 |
+
"nv": "product_20260921_nv-nv.webp"
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"name": "typography",
|
| 46 |
+
"seed": 20260920,
|
| 47 |
+
"prompt": "Design a polished square travel poster with a restrained mid-century screen-print illustration of a red tram crossing a stone bridge over a blue river at sunset. Large cream headline at the top reads exactly \"TAKE THE SCENIC ROUTE\". A smaller line at the bottom reads exactly \"SLOW TRAVEL CLUB\". Strong typographic hierarchy, generous margins, flat navy, coral and cream inks, subtle paper grain. All lettering sharp and clearly legible.",
|
| 48 |
+
"bf": "typography_20260920_bf-bf.webp",
|
| 49 |
+
"nv": "typography_20260920_nv-nv.webp"
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"name": "typography",
|
| 53 |
+
"seed": 20260921,
|
| 54 |
+
"prompt": "Design a polished square travel poster with a restrained mid-century screen-print illustration of a red tram crossing a stone bridge over a blue river at sunset. Large cream headline at the top reads exactly \"TAKE THE SCENIC ROUTE\". A smaller line at the bottom reads exactly \"SLOW TRAVEL CLUB\". Strong typographic hierarchy, generous margins, flat navy, coral and cream inks, subtle paper grain. All lettering sharp and clearly legible.",
|
| 55 |
+
"bf": "typography_20260921_bf-bf.webp",
|
| 56 |
+
"nv": "typography_20260921_nv-nv.webp"
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"name": "spatial",
|
| 60 |
+
"seed": 20260920,
|
| 61 |
+
"prompt": "A clean studio photograph on a light gray tabletop showing exactly three objects arranged in a straight horizontal row: a red ceramic cube on the left, a blue glass sphere in the center, and a yellow wooden cone on the right. The sphere is slightly larger than the cube, and the cone is tallest. A single soft light from the upper left casts shadows toward the lower right. Neutral pale gray background, eye-level camera, every object fully visible and separated by equal gaps, no other objects, no text.",
|
| 62 |
+
"bf": "spatial_20260920_bf-bf.webp",
|
| 63 |
+
"nv": "spatial_20260920_nv-nv.webp"
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"name": "spatial",
|
| 67 |
+
"seed": 20260921,
|
| 68 |
+
"prompt": "A clean studio photograph on a light gray tabletop showing exactly three objects arranged in a straight horizontal row: a red ceramic cube on the left, a blue glass sphere in the center, and a yellow wooden cone on the right. The sphere is slightly larger than the cube, and the cone is tallest. A single soft light from the upper left casts shadows toward the lower right. Neutral pale gray background, eye-level camera, every object fully visible and separated by equal gaps, no other objects, no text.",
|
| 69 |
+
"bf": "spatial_20260921_bf-bf.webp",
|
| 70 |
+
"nv": "spatial_20260921_nv-nv.webp"
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"name": "illustration",
|
| 74 |
+
"seed": 20260920,
|
| 75 |
+
"prompt": "A richly detailed storybook gouache illustration of a small fox librarian wearing round spectacles and a moss-green waistcoat, standing on a rolling wooden ladder in a towering circular library inside an ancient tree. The fox reaches toward a glowing blue book with one paw. Curving shelves, warm amber lamps, tiny floating dust motes, a rainy moonlit window, hand-painted brush texture, charming expressive face, coherent architecture, dramatic but cozy composition, no lettering.",
|
| 76 |
+
"bf": "illustration_20260920_bf-bf.webp",
|
| 77 |
+
"nv": "illustration_20260920_nv-nv.webp"
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"name": "illustration",
|
| 81 |
+
"seed": 20260921,
|
| 82 |
+
"prompt": "A richly detailed storybook gouache illustration of a small fox librarian wearing round spectacles and a moss-green waistcoat, standing on a rolling wooden ladder in a towering circular library inside an ancient tree. The fox reaches toward a glowing blue book with one paw. Curving shelves, warm amber lamps, tiny floating dust motes, a rainy moonlit window, hand-painted brush texture, charming expressive face, coherent architecture, dramatic but cozy composition, no lettering.",
|
| 83 |
+
"bf": "illustration_20260921_bf-bf.webp",
|
| 84 |
+
"nv": "illustration_20260921_nv-nv.webp"
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"name": "fashion",
|
| 88 |
+
"seed": 20260920,
|
| 89 |
+
"prompt": "A candid editorial photograph of an adult woman with short dark curly hair and freckles, wearing a cobalt-blue tailored wool coat over a cream turtleneck, standing outside a small Parisian cafe on a rainy evening. She glances back over her shoulder with a slight natural smile, holding a folded red umbrella by its wooden handle. Fine raindrops on the coat, realistic skin pores, wet pavement reflections, warm cafe windows against cool blue dusk, waist-up framing, 85mm lens, subtle film grain, believable unretouched photography.",
|
| 90 |
+
"bf": "fashion_20260920_bf-bf.webp",
|
| 91 |
+
"nv": "fashion_20260920_nv-nv.webp"
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"name": "fashion",
|
| 95 |
+
"seed": 20260921,
|
| 96 |
+
"prompt": "A candid editorial photograph of an adult woman with short dark curly hair and freckles, wearing a cobalt-blue tailored wool coat over a cream turtleneck, standing outside a small Parisian cafe on a rainy evening. She glances back over her shoulder with a slight natural smile, holding a folded red umbrella by its wooden handle. Fine raindrops on the coat, realistic skin pores, wet pavement reflections, warm cafe windows against cool blue dusk, waist-up framing, 85mm lens, subtle film grain, believable unretouched photography.",
|
| 97 |
+
"bf": "fashion_20260921_bf-bf.webp",
|
| 98 |
+
"nv": "fashion_20260921_nv-nv.webp"
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"name": "night_market",
|
| 102 |
+
"seed": 20260920,
|
| 103 |
+
"prompt": "A cinematic street photograph of a crowded night market in Taipei during gentle rain. In the foreground an elderly cook in a white apron lifts a steaming bamboo basket from a street-food stall, with hands clearly visible. A small illuminated sign above him reads exactly \"NIGHT BITES\". Rich red lanterns and teal shop lights reflect in wet asphalt; layered pedestrians with umbrellas recede into the scene. Realistic steam, appetizing food texture, natural human proportions, 35mm lens, documentary photography, atmospheric but restrained color.",
|
| 104 |
+
"bf": "night_market_20260920_bf-bf.webp",
|
| 105 |
+
"nv": "night_market_20260920_nv-nv.webp"
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"name": "night_market",
|
| 109 |
+
"seed": 20260921,
|
| 110 |
+
"prompt": "A cinematic street photograph of a crowded night market in Taipei during gentle rain. In the foreground an elderly cook in a white apron lifts a steaming bamboo basket from a street-food stall, with hands clearly visible. A small illuminated sign above him reads exactly \"NIGHT BITES\". Rich red lanterns and teal shop lights reflect in wet asphalt; layered pedestrians with umbrellas recede into the scene. Realistic steam, appetizing food texture, natural human proportions, 35mm lens, documentary photography, atmospheric but restrained color.",
|
| 111 |
+
"bf": "night_market_20260921_bf-bf.webp",
|
| 112 |
+
"nv": "night_market_20260921_nv-nv.webp"
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"name": "architecture",
|
| 116 |
+
"seed": 20260920,
|
| 117 |
+
"prompt": "Architectural interior photograph of a sunken conversation pit in a restored 1970s coastal house. Rust-orange corduroy sofas surround a low circular walnut table holding a clear glass vase with three white tulips. Floor-to-ceiling windows reveal a stormy ocean. Rain beads on the windows; a suspended black fireplace glows softly at the left. Pale travertine floor, detailed walnut ceiling, soft overcast daylight, balanced wide-angle composition, realistic materials and straight architectural lines, no people or lettering.",
|
| 118 |
+
"bf": "architecture_20260920_bf-bf.webp",
|
| 119 |
+
"nv": "architecture_20260920_nv-nv.webp"
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"name": "architecture",
|
| 123 |
+
"seed": 20260921,
|
| 124 |
+
"prompt": "Architectural interior photograph of a sunken conversation pit in a restored 1970s coastal house. Rust-orange corduroy sofas surround a low circular walnut table holding a clear glass vase with three white tulips. Floor-to-ceiling windows reveal a stormy ocean. Rain beads on the windows; a suspended black fireplace glows softly at the left. Pale travertine floor, detailed walnut ceiling, soft overcast daylight, balanced wide-angle composition, realistic materials and straight architectural lines, no people or lettering.",
|
| 125 |
+
"bf": "architecture_20260921_bf-bf.webp",
|
| 126 |
+
"nv": "architecture_20260921_nv-nv.webp"
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"name": "macro",
|
| 130 |
+
"seed": 20260920,
|
| 131 |
+
"prompt": "Extreme macro nature photograph of a tiny emerald jumping spider perched on the curled edge of a copper-colored autumn leaf. A single spherical dew drop beside the spider reflects a miniature garden. Fine individual hairs on the spider, realistic eight-legged anatomy, sharp jewel-like eyes, translucent dew, intricate leaf veins. Soft warm backlight, muted olive background with circular bokeh, narrow but carefully placed focus plane, natural colors, scientifically plausible wildlife photography, no lettering.",
|
| 132 |
+
"bf": "macro_20260920_bf-bf.webp",
|
| 133 |
+
"nv": "macro_20260920_nv-nv.webp"
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"name": "macro",
|
| 137 |
+
"seed": 20260921,
|
| 138 |
+
"prompt": "Extreme macro nature photograph of a tiny emerald jumping spider perched on the curled edge of a copper-colored autumn leaf. A single spherical dew drop beside the spider reflects a miniature garden. Fine individual hairs on the spider, realistic eight-legged anatomy, sharp jewel-like eyes, translucent dew, intricate leaf veins. Soft warm backlight, muted olive background with circular bokeh, narrow but carefully placed focus plane, natural colors, scientifically plausible wildlife photography, no lettering.",
|
| 139 |
+
"bf": "macro_20260921_bf-bf.webp",
|
| 140 |
+
"nv": "macro_20260921_nv-nv.webp"
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"name": "astronaut",
|
| 144 |
+
"seed": 20260920,
|
| 145 |
+
"prompt": "A cinematic close-up photograph of an adult female astronaut inside a weathered lunar habitat, looking directly into the camera through a clear helmet visor. Her brown eyes, small freckles and loose strands of hair remain clearly visible behind subtle reflections. White spacesuit with stitched fabric, orange fittings and a small rectangular chest patch reading \"LUNA 09\". Warm practical light on one cheek and cool Earthlight on the other, Earth visible through a round window in the background, restrained analog science-fiction realism, finely detailed materials, no illustration style.",
|
| 146 |
+
"bf": "astronaut_20260920_bf-bf.webp",
|
| 147 |
+
"nv": "astronaut_20260920_nv-nv.webp"
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"name": "astronaut",
|
| 151 |
+
"seed": 20260921,
|
| 152 |
+
"prompt": "A cinematic close-up photograph of an adult female astronaut inside a weathered lunar habitat, looking directly into the camera through a clear helmet visor. Her brown eyes, small freckles and loose strands of hair remain clearly visible behind subtle reflections. White spacesuit with stitched fabric, orange fittings and a small rectangular chest patch reading \"LUNA 09\". Warm practical light on one cheek and cool Earthlight on the other, Earth visible through a round window in the background, restrained analog science-fiction realism, finely detailed materials, no illustration style.",
|
| 153 |
+
"bf": "astronaut_20260921_bf-bf.webp",
|
| 154 |
+
"nv": "astronaut_20260921_nv-nv.webp"
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"name": "botanical_poster",
|
| 158 |
+
"seed": 20260920,
|
| 159 |
+
"prompt": "An elegant contemporary botanical exhibition poster on textured ivory paper. A lifelike ink-and-watercolor illustration of one flowering magnolia branch curves diagonally across the center, pale pink petals with delicate translucent edges and deep green leaves. At the top, a large refined serif headline reads exactly \"THE QUIET GARDEN\". Beneath it, smaller text reads exactly \"BOTANICAL STUDIES\". At the bottom, a small centered line reads exactly \"APRIL 12 - MAY 30\". Generous negative space, exceptional typography, restrained burgundy and sage palette, museum-quality printed design.",
|
| 160 |
+
"bf": "botanical_poster_20260920_bf-bf.webp",
|
| 161 |
+
"nv": "botanical_poster_20260920_nv-nv.webp"
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"name": "botanical_poster",
|
| 165 |
+
"seed": 20260921,
|
| 166 |
+
"prompt": "An elegant contemporary botanical exhibition poster on textured ivory paper. A lifelike ink-and-watercolor illustration of one flowering magnolia branch curves diagonally across the center, pale pink petals with delicate translucent edges and deep green leaves. At the top, a large refined serif headline reads exactly \"THE QUIET GARDEN\". Beneath it, smaller text reads exactly \"BOTANICAL STUDIES\". At the bottom, a small centered line reads exactly \"APRIL 12 - MAY 30\". Generous negative space, exceptional typography, restrained burgundy and sage palette, museum-quality printed design.",
|
| 167 |
+
"bf": "botanical_poster_20260921_bf-bf.webp",
|
| 168 |
+
"nv": "botanical_poster_20260921_nv-nv.webp"
|
| 169 |
+
}
|
| 170 |
+
]
|
assets/edit_portrait_bf-bf_extended_00001_.png
ADDED
|
Git LFS Details
|
assets/edit_portrait_nv-nv_extended_00001_.png
ADDED
|
Git LFS Details
|
assets/edit_product_bf-bf_extended_00001_.png
ADDED
|
Git LFS Details
|