pottokao/Qwen-Image-2.1-DiT-GGUF

Qwen-Image-2.1 DiT — GGUF (Q8_0 · Q6_K · Q4_K_M)

Not affiliated with, or endorsed by, Alibaba / Qwen. Quantized redistribution of the Qwen/Qwen-Image-2.1 diffusion transformer under the Qwen Research License (copy included as LICENSE). Non-commercial use only.

Loads in stock city96/ComfyUI-GGUF — no patch, no extra nodes. These files carry general.architecture = qwen_image, so UnetLoaderGGUF recognises them directly (GGUFs without that field fail with Unknown model architecture!).

Text encoder: our Qwen-Image-2.1 Text Encoder GGUF works with this DiT, but it needs one extra node first — ComfyUI-GGUF-Qwen3VL-TE (stock ComfyUI-GGUF doesn't load Qwen3-VL text encoders yet). With a safetensors text encoder (bf16 / int8 / FP8 from the same repo family) nothing extra is needed.

Made with a simple node

The sample / showcase images in this repo were generated with the technique behind ComfyUI-QwenImage-PhotoStyles — a small ComfyUI node (17 photographic styles; one short prompt becomes a full styled prompt via the PE-T2I rewriter). If you're interested in how the images are made, the node and its recipe are there.

Where are the prompts? The node is applied up front as the PE-T2I step — you type one short line and it expands it into the full prompt. The exact short prompts live in that repo's Gallery: the line printed under each image there is the literal prompt (copy it verbatim), not a description of the picture.

📥 Files

File	Size	For
qwen_image_2.1-Q8_0.gguf	7.69 GB	closest to bf16
qwen_image_2.1-Q6_K.gguf	6.00 GB	nearly identical to Q8_0 in our test
qwen_image_2.1-Q4_K_M.gguf	4.34 GB	recommended — smallest; small detail changes vs Q8_0

Put one in ComfyUI/models/diffusion_models/ (or models/unet/) and load it with UnetLoaderGGUF. Needs ComfyUI 0.36.0 or newer (for TextEncodeQwenImage21).

🖼️ Showcase

Warm Documentary	Candid B&W Street	Vivid Everyday Color
Black Fury	Geometry of Light	Frozen Film Still
Deadpan Large-Format	White Leap	Corner Elegance
Raw B&W Fashion	Color Behind Glass	High-Contrast Reportage
Lyrical Landscape	Hyper Chic East	Street in the Mirror
Plain Portrait	Melancholy B&W	Fairytale Giants
Cold Power	Cold Power	Hyper Chic East
Fairytale Giants	Hyper Chic East	White Leap

This is what Q4_K_M across the board looks like — every image above was made with:

Part	File
PE-T2I rewriter	pe_t2i_heretic-Q4_K_M.gguf
Text encoder	qwen3vl_8b_heretic-Q4_K_M.gguf + f16 mmproj
DiT	qwen_image_2.1-Q4_K_M.gguf
VAE	qwen_image_2.1_vae_bf16.safetensors (official)

_{AI-generated. Style homage only — not affiliated with or endorsed by the photographers.}

🍱 All-GGUF Qwen-Image-2.1

The whole pipeline can run on GGUF — useful on Macs and smaller GPUs:

Part	File	Where	Loader
Prompt rewriter (optional)	pe_t2i_heretic-Q4_K_M.gguf + system_prompt.txt	PE-T2I Heretic GGUF	llama.cpp / LLM node
Text encoder	qwen3vl_8b_heretic-Q4_K_M.gguf + mmproj-…-f16.gguf	Text Encoder Heretic GGUF	CLIPLoaderGGUF + patch node (until ComfyUI-GGUF supports Qwen3-VL)
DiT	qwen_image_2.1-Q4_K_M.gguf (or Q6_K / Q8_0)	this repo	UnetLoaderGGUF — stock
VAE	qwen_image_2.1_vae_bf16.safetensors (official)	—	VAELoader

Prefer no patch at all? Use a safetensors text encoder instead (e.g. bf16 or int8-convrot) with the stock CLIPLoader.

⚙️ Recommended sampler settings

25 steps, euler / simple. For images with text (signs, posters, labels), switch cfg partway through with two KSamplerAdvanced nodes — the first 15 steps at cfg 1 lock in composition and materials, the last 10 at cfg 3 with a negative prompt redraw the lettering:

	add_noise	steps	cfg	start_at_step	end_at_step	return_with_leftover_noise
1st	enable	25	1.0	0	12–17	enable
2nd	disable	25	3.0	12–17	25	disable

Example negative prompt: oversaturated, overexposed, gibberish text. Without text, a single KSampler at cfg 1.0 is fine.

Quantization

Standard ComfyUI-GGUF image-model rules — no per-layer sensitivity scan:

Tensors	Q4_K_M	Q6_K	Q8_0
img_in, txt_in, time_text_embed, norm_out, proj_out, modulation	BF16 (kept)	BF16 (kept)	BF16 (kept)
attn.to_v (×32 blocks)	Q6_K	Q6_K	Q8_0
attn.to_q / to_k / to_out, img_mlp.gate_up / out	Q4_K	Q6_K	Q8_0
norms	F32	F32	F32

How it was made (reproducible)

Converted from the official bf16 weights (ComfyUI layout, 265 tensors) with ComfyUI-GGUF's own tooling. Neither its convert.py nor its patched llama.cpp knew Qwen-Image, so both got a small addition — the patches are in tools/:

• tools/convert-qwen_image.patch — adds a qwen_image template to convert.py (7 lines)
• tools/llama.cpp-b3962-qwen_image.patch — ComfyUI-GGUF's lcpp.patch plus a qwen_image architecture and its keep-in-BF16 list

# ComfyUI-GGUF's tools/ with tools/convert-qwen_image.patch applied
python tools/convert.py --src qwen_image_2.1_bf16.safetensors --dst qwen_image_2.1-BF16.gguf
# llama.cpp at tag b3962 with tools/llama.cpp-b3962-qwen_image.patch applied
llama-quantize qwen_image_2.1-BF16.gguf qwen_image_2.1-Q4_K_M.gguf Q4_K_M   # also Q6_K, Q8_0

Verified

On ComfyUI 0.36.0 + ComfyUI-GGUF 6ea2651 (NVIDIA GPU):

• Stock loader — ComfyUI started with only ComfyUI-GGUF enabled (every other custom node disabled): all three files load through UnetLoaderGGUF with no compatibility-mode warning and generate normally. In that test (one prompt, one seed), they give near-identical images; Q6_K and Q8_0 are almost indistinguishable, Q4_K_M differs in small details. Rendered Chinese text (shop sign and a handwritten price menu) came out correct in all three.
• All-GGUF pipeline — Q4_K_M together with the Q4_K_M GGUF text encoder and the GGUF prompt rewriter.

Not tested on a Mac. Not tested: a numeric quality comparison against bf16.