Delete assets/resolution_sweeps/raw-LoRA-7steps-experiment

assets/resolution_sweeps/raw-LoRA-7steps-experiment/README.md

@@ -1,105 +0,0 @@
-# Krea 2 Raw at 7 steps with this LoRA — an experiment, not a supported setting
-
-**This LoRA is trained on Krea 2 Turbo, for Krea 2 Turbo, at 2 steps.** Nothing in its training saw Krea 2 Raw. It loads
-on Raw only because the two models share one architecture, so everything in this folder is an out-of-spec experiment
-rather than a way to use the adapter.
-
-These are the 15 test prompts, seed 4242, at 1024×768: **Krea 2 Raw + this LoRA at strength 1.0, run for 7 steps** — a
-quarter of Raw's usual 28 — with **light guidance** and an empty negative prompt.
-
-**Not every prompt works well at 7 steps, even with the LoRA.** Some come out as finished pictures, more or less:
-[portrait](1024x768/portrait.jpg), [kingfisher](1024x768/kingfisher.jpg), [market](1024x768/market.jpg),
-[pizza](1024x768/pizza.jpg), [mecha](1024x768/mecha.jpg), [storybook fox](1024x768/storybookfox.jpg),
-[colony ship](1024x768/colonyship.jpg), [snow leopard](1024x768/snowleopard.jpg), [inventor](1024x768/inventor.jpg) and
-[clay chef](1024x768/claychef.jpg). Others do not: [neon street](1024x768/neonstreet.jpg) is underexposed,
-[megacity](1024x768/megacity.jpg) is not as detailed as the Turbo version at 2 steps,
-[sorceress](1024x768/sorceress.jpg) is not the most aesthetic result, [race car](1024x768/racecar.jpg) lacks detail and
-[swordsman](1024x768/swordsman.jpg) has less detail than the Turbo version at 2 steps — and for those, a few more steps
-may help. In the other direction, some prompts give a usable quick preview at even 4–5 steps: not as good as 7 steps or
-more, but not broken either. Looked at closely: small and distant subjects hold up better than you might expect, with
-the mid-distance faces in the [market](1024x768/market.jpg) crowd distinct rather than melted, while the drone in
-[megacity](1024x768/megacity.jpg) is only a smeared shape. Treat this as an experiment and look at each image on its own
-rather than as a promise of what Raw at 7 steps will give.
-
-**More or less finished at 7 steps** — click any image for full size:
-
-[![portrait — Krea 2 Raw + this LoRA, 7 steps](thumbs/portrait.jpg)](1024x768/portrait.jpg)
-[![kingfisher — Krea 2 Raw + this LoRA, 7 steps](thumbs/kingfisher.jpg)](1024x768/kingfisher.jpg)
-[![market — Krea 2 Raw + this LoRA, 7 steps](thumbs/market.jpg)](1024x768/market.jpg)
-[![pizza — Krea 2 Raw + this LoRA, 7 steps](thumbs/pizza.jpg)](1024x768/pizza.jpg)
-[![mecha — Krea 2 Raw + this LoRA, 7 steps](thumbs/mecha.jpg)](1024x768/mecha.jpg)
-[![storybook fox — Krea 2 Raw + this LoRA, 7 steps](thumbs/storybookfox.jpg)](1024x768/storybookfox.jpg)
-[![colony ship — Krea 2 Raw + this LoRA, 7 steps](thumbs/colonyship.jpg)](1024x768/colonyship.jpg)
-[![snow leopard — Krea 2 Raw + this LoRA, 7 steps](thumbs/snowleopard.jpg)](1024x768/snowleopard.jpg)
-[![inventor — Krea 2 Raw + this LoRA, 7 steps](thumbs/inventor.jpg)](1024x768/inventor.jpg)
-[![clay chef — Krea 2 Raw + this LoRA, 7 steps](thumbs/claychef.jpg)](1024x768/claychef.jpg)
-
-**Weaker at 7 steps** — more steps may help these:
-
-[![neon street — underexposed — Krea 2 Raw + this LoRA, 7 steps](thumbs/neonstreet.jpg)](1024x768/neonstreet.jpg)
-[![megacity — less detail than Turbo at 2 steps — Krea 2 Raw + this LoRA, 7 steps](thumbs/megacity.jpg)](1024x768/megacity.jpg)
-[![sorceress — not the most aesthetic — Krea 2 Raw + this LoRA, 7 steps](thumbs/sorceress.jpg)](1024x768/sorceress.jpg)
-[![race car — lacks detail — Krea 2 Raw + this LoRA, 7 steps](thumbs/racecar.jpg)](1024x768/racecar.jpg)
-[![swordsman — less detail than Turbo at 2 steps — Krea 2 Raw + this LoRA, 7 steps](thumbs/swordsman.jpg)](1024x768/swordsman.jpg)
-
-The guidance matters. At 7 steps, guidance 4.5 crushed most of these prompts into near-black frames with clipped colour,
-with or without the LoRA, and on the portrait we tried with no guidance at all, Raw on its own came out washed out and
-ghosted. Light guidance with the LoRA is the combination that turned 7 steps into pictures in our tests. The same
-prompts rendered by stock Raw at the same 7 steps and light guidance, without the LoRA, are in the subfolder
-[`_raw-base-NO-LoRA-7step-cfg1/`](_raw-base-NO-LoRA-7step-cfg1), and side by side with this LoRA's images in the [2 way
-RAW comparison](#2-way-raw-comparison) below.
-
-## 2 way RAW comparison
-
-Each strip is one prompt with the same seed at 1024×768, 7 steps, light guidance (1.0) and an empty negative prompt,
-rendered twice: stock Krea 2 Raw on the left, Raw + this LoRA on the right, so the LoRA is the only difference. In
-ComfyUI terms, guidance 1.0 is cfg 2.0. Click any strip for full size:
-
-[![portrait — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/portrait_2way.jpg)](comparison/portrait_2way.jpg)
-[![kingfisher — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/kingfisher_2way.jpg)](comparison/kingfisher_2way.jpg)
-[![market — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/market_2way.jpg)](comparison/market_2way.jpg)
-[![pizza — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/pizza_2way.jpg)](comparison/pizza_2way.jpg)
-[![mecha — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/mecha_2way.jpg)](comparison/mecha_2way.jpg)
-[![storybook fox — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/storybookfox_2way.jpg)](comparison/storybookfox_2way.jpg)
-[![colony ship — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/colonyship_2way.jpg)](comparison/colonyship_2way.jpg)
-[![snow leopard — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/snowleopard_2way.jpg)](comparison/snowleopard_2way.jpg)
-[![inventor — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/inventor_2way.jpg)](comparison/inventor_2way.jpg)
-[![clay chef — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/claychef_2way.jpg)](comparison/claychef_2way.jpg)
-[![neon street — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/neonstreet_2way.jpg)](comparison/neonstreet_2way.jpg)
-[![megacity — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/megacity_2way.jpg)](comparison/megacity_2way.jpg)
-[![sorceress — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/sorceress_2way.jpg)](comparison/sorceress_2way.jpg)
-[![race car — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/racecar_2way.jpg)](comparison/racecar_2way.jpg)
-[![swordsman — stock Raw · Raw + this LoRA, 7 steps, guidance 1.0](thumbs/swordsman_2way.jpg)](comparison/swordsman_2way.jpg)
-
-## ComfyUI
-
-[`krea2_raw_7step_lora_experiment_t2i.json`](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-2step-LoRA/blob/main/krea2_raw_7step_lora_experiment_t2i.json),
-in the root of this repository, reproduces the portrait experiment in ComfyUI with stock nodes only: Krea 2 Raw
-(`krea2_raw_bf16.safetensors` from [Comfy-Org/Krea-2](https://huggingface.co/Comfy-Org/Krea-2)) with this LoRA's
-`_comfyui` file at strength 1.0, **7 steps**, **cfg 2.0**, sampler euler, scheduler simple, 1024×768, seed 4242 and an
-empty negative prompt. Seeds do not carry over between runtimes, so it will not reproduce the images in this folder
-pixel for pixel.
-
-[![ComfyUI workflow — Krea 2 Raw + this LoRA, 7 steps](thumbs/workflow_preview_raw.jpg)](workflow_preview_raw.jpg)
-
-The main workflow next to it, `krea2_turbo_2step_lora_t2i.json`, can do the same. Load it and change four
-things; the LoRA node and its `_comfyui` file stay as they are:
-
-1. **Load Diffusion Model** → `krea2_raw_bf16.safetensors` (or `krea2_raw_int8_convrot`; `krea2_raw_fp8_scaled` is
-   NVIDIA-only)
-2. **KSampler** → steps **7**, cfg **2.0**
-3. **the negative** → replace the `ConditioningZeroOut` with a **CLIP Text Encode** left empty: at cfg 2.0 the negative is
-   evaluated, and a zeroed conditioning is not the same as an empty prompt
-4. **Empty Latent Image** → 1024 × 768, or any size you want to try
-
-On cfg: cfg 1.0 in ComfyUI turns guidance off, and a heavy cfg such as 5.5 (guidance 4.5) crushes most prompts at 7
-steps; the light guidance these images use is cfg 2.0.
-
-Layout: `1024x768/<prompt>.jpg`, one image per prompt with this LoRA, the same prompts and naming as the other trees
-here; `_raw-base-NO-LoRA-7step-cfg1/1024x768/` holds the same prompts on stock Raw at the same settings;
-`comparison/<prompt>_2way.jpg` holds the 2 way strips; `thumbs/` holds the reduced copies shown above (`<prompt>.jpg`,
-`<prompt>_2way.jpg` and `workflow_preview_raw.jpg`); `workflow_preview_raw.jpg` is a screenshot of the ComfyUI workflow,
-which sits in the root of this repository.
-
-For the settings the LoRA is distilled for — **Krea 2 Turbo, 2 steps, guidance 0.0 (cfg 1.0 in ComfyUI), mu = 1.15,
-strength 1.0** — see the [main README](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-2step-LoRA) at the root of this repository.