A curated list of checkpoints, quants, prompt engines, LoRAs, and tooling for Qwen-Image 2.1 — Alibaba's 7B unified text-to-image and image-editing model.

Pick your entry point based on the runtime you already have.

Official resources

• Qwen-Image-2.1 model card — official weights, license, aspect-ratio table
• Qwen-Image-2.1 GitHub repo — inference code, news, supported frameworks
• Qwen blog: Qwen-Image-2.1 — release notes with architecture and benchmark figures
• ComfyUI native workflow example — official ComfyUI tutorial
• ModelScope mirror — official CN mirror

Model facts worth knowing

• 7B parameters in the visual generation component, 32 single-stream DiT layers.
• Mixed-granularity attention (token-level causal for text, chunk-level for image) with prefix KV cache reuse for multi-reference editing.
• Native RGBA output — the prompt decides whether the result has an alpha channel.
• Editing accepts up to 10 reference images, plus circles, painted annotations, or separate masks for local edits.
• Qwen3-VL-8B is the text encoder, so the text encoder is the largest single download at ~17.5 GB BF16.
• Default sample setting is 40 steps without classifier-free guidance; CFG is available for prompt adherence.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

The reference weights. Start here before touching any community conversion.

· · · · · · · · · · · · · ·

Name	Precision	Layout	DiT	Text encoder	VAE	Links
Qwen-Image-2.1		diffusers	14.23 GB	17.53 GB (Qwen3-VL-8B)	1.35 GB

The reference release, and the only repo you need for a standard Diffusers setup. The DiT and text encoder ship as numbered safetensors shards with an index, so pull the repo rather than a single file.

· · · · · · · · · · · · · ·

Comfy-Org — the Day-0 ComfyUI repackage. Files land directly in models/ with no renaming.

Name	Precision	Size	Links
Image Model		14.23 GB
Image Model		7.26 GB
Text Encoder		17.53 GB
Text Encoder		9.35 GB
Text Encoder		6.31 GB
Prompt Engine T2I		9.47 GB
Prompt Engine I2I		9.47 GB
VAE		0.68 GB
ControlNet Union		7.55 GB
ControlNet Union		3.78 GB

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

Qwen-Image 2.1 conditions on Qwen3-VL-8B, which at BF16 is the largest file in the stack. It also ships dedicated prompt-rewriting models — separate text models that take a short request in any language and return a detailed English prompt plus a recommended aspect ratio. Neither is part of the diffusion path; both are optional, and you can run the base model without either.

· · · · · · · · · · · · · ·

The official pair is fine-tuned Qwen3.5-VL 9B. The Pocket builds are fine-tuned from Qwen3.5 base models instead — roughly 5× smaller, text-to-image only.

Name	Task	Params	Precision	Size	Links
PE-T2I		9B		18.82 GB
PE-I2I		9B		18.82 GB
PE-T2I Pocket 2B		2B		3.76 GB
PE-T2I Pocket 0.8B		0.8B		1.50 GB

The official rewriters ship a system_prompt.txt and work with AutoModelForCausalLM + AutoTokenizer. Output is JSON after a reasoning block, so split on the think-tag before parsing. The T2I rewriter also returns a recommended wh_ratio you can map straight to a size.

Each official repo splits into four model-0000N.safetensors shards plus a model.safetensors.index.json, so download the repo — there is no single-file build to link to. The 0.8B Pocket build also ships a Q8_0 GGUF (0.81 GB) for llama.cpp.

· · · · · · · · · · · · · ·

Type	Name	Task	Precision	Size	Links
	Heretic TE GGUF			33.07 GB
	Heretic TE DHQ GGUF			4.56 GB
	Heretic TE NVFP4			6.31 GB
	Heretic TE int8 ConvRot			9.35 GB
	Heretic TE W4A8			6.31 GB
	Heretic TE BF16			17.53 GB
	PE-T2I Heretic GGUF			5.89 GB
	PE-I2I Heretic GGUF			6.81 GB
	ComfyUI PE bundle			18.07 GB
	PE-I2I Abliterated			18.82 GB
	PE-T2I Heretic			18.82 GB
	PE-I2I Heretic			18.82 GB
	PE-T2I Heretic NVFP4			11.20 GB
	PE-I2I Heretic NVFP4			11.20 GB

Start with the TE GGUF build if you want one download: Q4_K_M (5.03 GB), fp8 (9.34 GB), bf16 (17.53 GB), plus a 1.16 GB mmproj projector, all in one repo.

DHQ (Dual-Head Quant) is a sibling of that build, not a different model: the same heretic weights re-quantized so each tensor gets the precision that matters to what the Qwen 2.1 DiT actually reads — measured through the DiT's own text-input stage rather than by next-token loss. Every file is byte-for-byte the same size as llama.cpp's build of that name and better at it. The ladder runs IQ2_S → IQ4_XS plus Q2_K / Q3_K_M / Q4_K_M, and each one ships its imatrix recipe. Take DHQ below Q5; at Q5 and above the standard recipe already matches it, so there is no reason to. It is tuned for this DiT specifically — Ideogram 4 and the other Qwen3-VL-8B consumers can load the files but were not tuned for.

· · · · · · · · · · · · · ·

Name	Format	Precision	Size	Links
PE ComfyUI pack	BF16 + int8 ConvRot		64.68 GB
PE-T2I MLX	MLX (4/8/16-bit)		35.20 GB
PE-I2I MLX	MLX (4/8/16-bit)		35.20 GB
Prompt Enhancement INT8	int8 ConvRot		24.69 GB
Heretic T2I int8 tensorwise	int8 tensorwise ConvRot		9.99 GB
Heretic PE int8 ConvRot	int8 ConvRot (T2I + I2I)		19.91 GB

The INT8 ConvRot pack covers both PE-T2I and PE-I2I in one download. The heretic int8 ConvRot pair does too, and unlike the tensorwise build it keeps the MTP head — 1,395 tensors against 829, one file per task at 9.96 GB.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

Community conversions of the base DiT. Sizes are total weight bytes per repo. Distilled turbo GGUFs live under Turbo instead.

· · · · · · · · · · · · · ·

Transformer-only weights for llama.cpp, sorted from the highest quant down. Q4_K_M is the usual quality/size balance point. Unsloth is the primary source — it carries the widest ladder, so it wins every quant it ships. Where two repos offer a quant Unsloth does not, both are linked in the same cell.

Quant	Size	Download
	14.23 GB
	14.23 GB
	7.64 GB
	6.72 GB
	6.27 GB
	5.39 GB
	4.50 GB
	5.96 GB
	4.05 GB
	4.20 GB
	4.15 GB	┊
	3.91 GB
	3.61 GB
	3.17 GB
	2.72 GB
	2.47 GB
Text encoder	6.30 GB
Text encoder	5.03 GB
mmproj projector	0.75 GB
VAE	0.68 GB
VAE	0.68 GB

· · · · · · · · · · · · · ·

Every FP4 / NVFP4 / MXFP4 / INT8 / INT4 / W4A4 conversion in one place.

Name	Precision	Size	Links	Notes
INT4ConvRot-ComfyUI		65.91 GB		Best all-in-one ComfyUI pack. DiT and TE at three precisions plus VAE, in correct folder layout.
Darkstar ModelOpt W4A4 NVFP4		23.66 GB		NVIDIA ModelOpt-derived, full repo.
INT8		17.96 GB		Full diffusers repo.
DiT NVFP4 ComfyUI		13.82 GB		Three ComfyUI cuts: nvfp4, nvfp4_T2, nvfp4_T3.
FP4		11.74 GB		DiT 6.46 + TE 4.94 + VAE 0.34.
INT4		11.08 GB		Full diffusers repo.
bnb 4bit		11.41 GB		Standard bitsandbytes NF4.
MXFP4 (Paiton)		9.31 GB		Paiton backend, 57 shards.
MXFP4 Paiton RDNA4		9.31 GB		RDNA4-specific kernel variant, identical layout.
Uncensored MXFP4 Paiton		9.31 GB	⚠️	Uncensored, derived from the abenzerps GGUF.
Noct Q Anime		7.26 GB	⚠️	Merged anime fine-tune of the DiT, single file, int8 ConvRot. Fits 8–12 GB cards. Without the word "anime" in the prompt you get a photo — the card is blunt that there is no trigger word. 25 steps, euler/simple, CFG 3. Uncensored: it will render explicit content. The card names no second parent model or merge recipe, so the lineage is Qwen 2.1 plus undisclosed transformer edits.
W4A4 NVFP4		4.88 GB		DiT only; near-identical to the INT4 build.
W4A4 INT4		4.66 GB		DiT only.
Uncensored NVFP4		4.20 GB	⚠️	DiT only, native NVFP4 rather than a GGUF conversion. Repo's own 192 F8_E4M3 scales against 192 FP32 originals.

· · · · · · · · · · · · · ·

Name	Precision	Size	Links	Notes
Darkstar ModelOpt FP8		26.33 GB		NVIDIA ModelOpt-derived, full repo.
FP8		17.96 GB		Closest thing to a drop-in smaller BF16.
Uncensored BF16 SafeTensor		14.23 GB	⚠️ ┊	Single 14.23 GB file — the full DiT in one piece, mirrored by both repos.
Uncensored Genesis BF16		32.46 GB	⚠️	The abenzerps uncensored DiT with the author's "Genesis" denoising pass applied — a post-training SVD repair, not a fine-tune, so it is a distinct weight set rather than a re-upload. Self-contained: DiT 14.23 + a Genesis Qwen3-VL-8B text encoder (16.39) + mmproj (1.16) + VAE 0.68.
DF11 ComfyUI		9.72 GB		qwen_image_2.1_bf16-DF11.safetensors.

· · · · · · · · · · · · · ·

SVDQ-based 4-bit for Nunchaku, which targets low-VRAM systems and 4090-class cards.

Name	Precision	Size	Links	Notes
nunchaku		8.75 GB		Two builds: best_quality_fp4 and svdq-fp4_r32.
nunchaku lite int4		4.15 GB		Half the size of the FP4 build.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

Few-step students of the base model, distilled with Distribution Matching Distillation. They trade fidelity for speed: useful for iteration and batch work, weaker than 40 steps on multi-reference composition and identity-preserving edits.

Current Viggle release: v0.2.1 (2026-09-24), 6 steps. It supersedes v0.2 (5-step name, sampled at 6) and v0.1 (4 steps). Sample with sigmas=[1.0, 0.9375, 0.875, 0.75, 0.5, 0.25], CFG 1.0, empty negative prompt. Official source: Viggle/Qwen-Image-2.1-viggle-turbo.

Three independent distillation lines are listed below, and their checkpoints are not interchangeable — pick one family. Viggle is the reference DMD line above. Alibaba PAI ships an official 4-step line via Parallel Decoding Distillation (PDD) in VideoX-Fun. Pruna is a third-party DMD line with 8-step and 5-step adapters.

The Pruna adapters are strict about sampling. Each has its own sigma schedule and they are not interchangeable, so load exactly one: 8-step wants sigmas=[1.0, 14/15, 6/7, 10/13, 2/3, 6/11, 0.4, 2/9], 5-step wants sigmas=[1.0, 0.94, 6/7, 2/3, 0.4]. Both need shift=1.0 with dynamic shifting off (otherwise the sigmas get shifted twice), no CFG, no negative prompt, and LoRA strength 1.0. Trained at 1K with up to 3 reference images; 2K runs but sits outside training coverage. Requires a pinned diffusers commit.

· · · · · · · · · · · · · ·

DiT-only quants for ComfyUI-GGUF, from the v0.1 full fine-tune (4-step). Both repos keep FP32 attention norms and BF16 patch embeddings, so do not stack a Viggle LoRA on top. Repos: realrebelai (wider ladder) and Abiray (adds Q4_K_S). Where both ship a quant, both are linked; sizes differ because the builds differ.

Quant	Size	Download
	7.69 / 7.59 GB	┊
	6.91 / 5.88 GB	┊
	5.96 / 5.01 GB	┊
	5.56 / 4.19 GB	┊
	4.06 GB
	4.19 / 3.19 GB	┊
	3.77 GB

· · · · · · · · · · · · · ·

Repo links: Viggle · alibaba-pai · Pruna · chfm (v0.2 snapshot) · t8star · xingewh · RunningHubAI · addlabsviral · cgb

Two ComfyUI conversions of the v0.2.1 adapters exist and are not interchangeable. xingewh re-keys the weights for ComfyUI and loads with the stock Load LoRA node. t8star fuses gate/up for ComfyUI's merged MLP, which is why its files are larger; load those with the built-in LoraLoaderBypassModelOnly at strength 1.0 on a BF16 base, because ordinary merging loaders lose adapter updates to rounding.

Name	Steps	Precision	Size	Links	Notes
v0.2.1 LoRA r256	6		1.36 / 1.36 / 1.76 GB	┊ ┊	Start here. Sharpest and most faithful to the 40-step base; runs the demo Space. Load on the base transformer at runtime — do not merge.
v0.2.1 LoRA r128	6		0.68 / 0.68 / 0.88 GB	┊ ┊	Same adapter cut to rank 128; what the shipped ComfyUI workflows use.
v0.2.1 adapter (peft)	6		2.72 GB		The v0.2.1 LoRA in PEFT key format, F32 as trained.
v0.2 LoRA r256	5 / 6		1.36 GB	┊ ┊	Step 600 of the same run. The 5step in the name is the launch schedule; sample at 6 like v0.2.1.
v0.2 LoRA r128	5 / 6		0.68 GB		Rank-128 cut of v0.2.
v0.1 full fine-tune	4		14.23 GB		Merged transformer, no LoRA needed. This is what the GGUF quants above were built from.
v0.1 LoRA r64	4		0.34 GB	┊ ┊	Superseded — diversity collapsed to 0.75× base and the output was visibly softer. Kept for reproducibility.
Fun-Acc 4Step (PDD)	4		0.35 GB		Official Alibaba PAI line, rank 64, via Parallel Decoding Distillation. Independent of Viggle — 4 NFE for both T2I and instruction editing.
Pruna 8Step	8		0.34 GB	┊	Third-party DMD line, rank 64 / alpha 128. Better of the two and the card's default, but v0.1 is marked work in progress.
Pruna 5Step	5		0.34 GB	┊	Same line, faster and visibly weaker. One adapter at a time — see the schedule note above.
UltraFast 8Step	8		0.17 GB		Editing-focused rather than T2I, trained on MagicBrush. Rank 32, but loaded by the repo's own inject_dit_lora script, not a stock PEFT or ComfyUI loader.
Turbo BF16 diffusers	4		32.44 GB		Full pipeline (TE + DiT + VAE), ready to load with QwenImage21Pipeline.
Turbo FP4 diffusers	4		11.41 GB		Same v0.1 pipeline with an FP4 DiT; the TE is the larger half at 6.73 GB.
Turbo ONNX (browser)	4		~17.2 GB		r64 LoRA merged into the denoiser, then Q4 MatMulNBits. WebGPU in-browser; needs the FreeGen pipeline and a desktop adapter. Experimental.
6NFE Distill	6		28.46 GB		Full transformer distill in 8 shards, 297 tensors FP32 — no LoRA to load. Handles T2I and editing. Needs shift_terminal=0.4 and conditional-only sampling (true_cfg_scale=1.0, no negative prompt) while keeping the base's dynamic shift, because the sixth evaluated sigma is 0.4. Ships its own inference.py; --cpu-offload trades speed for VRAM. Third-party, same research license.
UltraFast Q4_K	8		4.05 GB		DiT-only Q4_K with the UltraFast adapter already merged — do not stack the adapter on top. Still needs the text/vision encoders and the matching VAE. Quality and speed unbenchmarked.
Pruna 8Step SDNQ	8		11.51 GB		The Pruna 8-step adapter merged into the base and re-quantized with SDNQ — UINT4 dynamic, Hadamard rotation at group size 256. Full diffusers pipeline (DiT 4.10 + TE 6.74 + VAE 0.68), so the text encoder is quantized too. Needs SDNQ 0.2.0+, Triton and diffusers from git. Follows the Pruna sigma schedule above.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

Style, control, and fix adapters. All target the base DiT unless noted. Grouped by type, then by size.

Name	Type	Precision	Size	Links	Notes
ControlNet-Union			7.55 GB	┊	Official Alibaba PAI / VideoX-Fun. One checkpoint for 8 conditions (Canny, Depth, Grayscale, HED, Lineart, MLSD, Pose, Scribble) plus inpainting. Control branch only, 16 injection points, loaded strict=False. The t8star link is the same weights re-saved under ComfyUI naming, for 368 bytes more. ComfyUI users can skip both and take the model_patches/ cut from Comfy-Org above, which also has an int8 build at 3.78 GB.
Object Mover Bbox Preview			0.50 GB		Bbox object moving, 6 checkpoints.
Object Remover Bbox Preview			0.50 GB		Bbox object removal, full-quality variant.
Object Remover Bbox turbo			0.42 GB		4-step-compatible variant.
BFS Head Swap v1			0.32 GB		Head replacement, not a whole-face blend: identity, hair, eye colour and nose come from image 2 while gaze direction, head rotation and expression stay with image 1. Rank 64, 5,000 steps, MIT. Image order is load-bearing — swapping the two inputs swaps who is retargeted. Trigger with the head_swap: prefix. The repo's other 17 adapters target Qwen Image Edit 2509/2511, Flux 2 Klein, Krea 2 and LTX-2, not this base.
Orbit Alpha			0.17 GB		Official ML-Intern-lab. One RGBA image in, the same object from a new viewpoint out. Rank 32, 2,000 steps at 768 px. Use the _gate_up_split file — the other checkpoint in the repo does not load correctly. <orbit> grammar, 40 steps, no CFG.
FG extract v4			0.17 GB		Cuts a masked element out as an RGBA cutout at the input's exact size and position — it extracts rather than redraws. Built on the Viggle turbo base, so 4 steps, no CFG; ~2 s at 512², ~9 s at 1024² with cleaner edges. You supply the mask (SAM, box, or a detector): image 1 is the original dimmed to 20% outside the mask, image 2 is the plain RGB original. Do not pass the mask itself as an input image or it gets copied into the result. Needs an alpha-capable VAE for transparency. v3 is the fixed-prompt fallback.
Doodle-in			0.17 GB		Turns a magenta scribble into the object you name, keeping its shape and pose. Diffusers users want the linked _gate_up_split file — ai-toolkit's fused img_mlp.gate_up is silently dropped otherwise, loading 384 of 448 modules. In ComfyUI use doodle_in_lora_qwen21.safetensors instead. Marker is pure magenta RGB (255, 0, 255) at 3–7 px.
Outpaint v2			0.16 GB		Pad the picture with flat #808080, hand the padded canvas to the model as the reference; the adapter fills the gray and keeps the original pixel-registered. One side, a corner, or all four. Rank 32, ComfyUI keys, 2,000 steps. 25 steps, CFG 1, resolution 0, target 1–2 MP. Do not pin the known area with a latent noise mask — on this model it draws a visible rectangle at the seam.
Outpaint v1			0.16 GB		Same adapter one step earlier, trained at ≤1 MP over more extreme zoom-outs. The better of the two on very large extensions; the two are within noise on ordinary crops. The repo also keeps the step-500 and step-1250 intermediates.
AnyAngle			0.12 GB		Arbitrary camera-angle change, not the fixed azimuth/elevation of Orbit Alpha. Image 1 is the photo whose content and style you keep; image 2 is a coarse render from the angle you want, typically shot from a Gaussian splat or 3D model in Blender. Prompt is literally "Change the camera angle from <image2> to <image1>". Strength 1, CFG 3.0, 20+ steps — the only adapter here that wants CFG above 1. Quality is bounded by the splat: a bad reconstruction puts objects in the wrong place.
Anime consistency			0.17 GB		Character consistency while restyling anime images, trained from 4-view reference sheets. The author marks it experimental — start LoRA weight at 0.6–0.8. Diffusers keys with a transformer_blocks. prefix, no diffusion_model..
Edit consistency			0.16 GB		Stops edits from resizing or nudging the picture — the same drift Outpaint fixes, from the opposite side. No trigger word, so it composes with any edit prompt. The card measures worst-corner drift falling from 24.3 px to 1.6 px on restyles. Rank 32. The linked step-1500 file keeps Qwen's own look; a step-2000 file in the repo aligns tighter but washes colour out.
Fix			0.11 GB		The most-liked community LoRA.
De-AI LoRA pack			2.45 GB		8-file "remove the AI look" pack (CN filenames).
Natural Exposure LoRA			0.42 GB		Exposure correction, 5 checkpoints.
Product relight			0.17 GB		Lights a product or object so it sits in its background: matching speculars, front-to-back falloff, a real contact shadow. Trigger pengyu at the very start of the prompt or it does nothing. The card documents no sampler or strength range, so treat the defaults as untested.
De-AI + lighting v5			0.17 GB		Two revisions of the same de-AI + lighting adapter; v5 is the newer. The repo also has two lighting-only adapters, 0.24 and 0.09 GB. CN filenames.
ZZZ style			0.16 GB		Zenless Zone Zero-inspired look, rank 32. Strength 0.8 is the card's sweet spot; 0.6–0.8 for lighter influence, above 1.0 it starts overriding your composition. 25 steps, CFG 4.3, euler/simple, shift 3.1. Nine checkpoints across v1.4 and v1.6 — take v1.4_step3000 or v1.4_step2000; step 1000 is under-trained and step 4000 over-fitted. Trained mostly on single-girl portraits, so multi-character and male subjects generalize poorly.
Sts2 Cards Drawer			0.10 GB		deckbuilder_cardart_style_lora_v1_fp16.
ZZZ chibi			0.08 GB		Compresses proportions to chibi rather than just restyling. Rank 16, F16, all 32 layers. Take st2500 at strength 0.70; st2000 is gentler, st3000 flattens clothing detail. Trained on the Comfy-Org int8 ConvRot base, so expect slightly less fine detail than an FP16-trained LoRA. Separate from the ZZZ style row — not a stacked pair.
Realism (Nano Banana)			0.004 GB		Makes a generated image read as a real phone photo — skin texture, single-source lighting, sensor noise. LoKR, not LoRA: needs a LyCORIS loader, and it is the only entry in this table that does. Checkpoints at 500 / 750 / 1000. Use strength 0.5–0.7, not 1.0 — the card is explicit that 1.0 is heavy.
Normal2RGB			0.25 GB		Normal map → RGB render, 3 checkpoints.
Hips / buttocks			0.47 GB	⚠️	Body-shape slider, ported from a Qwen-2.5-09 derivative. CN filename.
Breasts and hips			0.17 GB	⚠️	Combined body-shape adapter from the same pack. CN filename.
RadianceChrome Voluptuous			0.17 GB	⚠️	Character-style LoRA.
NSFW LoRA			0.16 GB	⚠️	Civitai original by TheseAlpacas, mirrored unmodified. Rank 32, 192 targets. The author asks for 25+ steps, er_sde and the beta scheduler on an INT8 ConvRot base — it is not a few-step adapter.
NSFW Image Edit			0.08 GB	⚠️	Editing LoRA, uploaded 2026-09-24.
Uncensored LoRA			0.03 GB	⚠️	The adapter form of the uncensored weights, from the most-downloaded repo here. Undocumented on the card — the full weights are already in that repo's GGUF ladder, so reach for this only if you want the uncensored behaviour without replacing the DiT.
Breasts Slider V1			0.003 GB	⚠️	Slider control, 3 MB.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

Non-CUDA runtimes and specialized accelerator backends.

· · · · · · · · · · · · · ·

Name	Format	Precision	Size	Links	Notes
MLX 8bit	MLX (mflux)		24.04 GB		For mflux. 13 TE shards.
Coreml	CoreML .mlpackage		14.74 GB		4 transformer blocks (~3.5 GB each) + embed + VAE decoder.
QIPACK base	QIPACK1 .qipack		14.23 GB		For the native C++/Metal qwen-image-cplus runtime. 40-step base, defaults to TaylorSeer caching.
QIPACK distilled	QIPACK1 .qipack		14.23 GB		Same runtime, 4-step. The Viggle v0.1 full fine-tune, not its LoRA and not v0.2.1.
MLX 4bit	MLX		11.59 GB		Smallest viable Apple build.
Uncensored MLX	MLX (4/6/8-bit)		7.56 GB	⚠️	The uncensored line's Apple build, from the GGUF repo. DiT only — 7.56 GB at 8-bit, 5.78 at 6-bit, 4.00 at 4-bit, and the repo ships no MLX text encoder or config, so pair it with one of the two repos above.

The repo is self-contained: alongside the two packs it now ships the Qwen3-VL-8B text encoder (4 shards, 17.53 GB), the VAE (1.35 GB) and the processor files, so the support download is no longer needed. Needs macOS 14+. 1024×1024 works; 2048×2048 does not yet.

· · · · · · · · · · · · · ·

Alibaba MNN runtime for on-device inference. The full repos are large — the MNN build bundles the text encoder, DiT, and VAE together.

Name	Precision	Size	Links	Notes
MNN fp16		30.72 GB		Highest fidelity, and by far the largest.
MNN int8		21.42 GB
MNN int4		14.39 GB
MNN		10.62 GB		llm.mnn.weight 4.73 + dit.mnn.weight 4.47 + VAE 0.51 GB.

· · · · · · · · · · · · · ·

p150 is the standout entry here. A full port to a single Tenstorrent Blackhole p150a via tt-nn, with all three sub-models resident on-chip. ~20.5 s for a 40-step 1024² generation (513 ms/step sustained), and it beats an RTX 5090 with CPU offload by 1.4–1.6× end to end. Includes editing support for 1–4 condition images. No weights are redistributed — it pulls the official ones.

ROCm gfx1151 targets AMD Strix Halo / RX 9070-class iGPUs. Code only; no weights in the repo.

Intel Arc (SYCL). Frosty40 is a serving package for the sd.cpp SYCL backend, not a fine-tune — "turbo" refers to the recipe, not to step distillation. Despite the name the weights are Qwen's own, quantized from a BF16 master.

Name	Precision	Size	Links	Notes
SYCL quality		7.89 GB		Attention qkv/o kept at F16; 73 BF16 + 128 F16 + 96 Q5_0.
SYCL lean		5.07 GB		All-exception Q5_0, 73 BF16 + 224 Q5_0.

Both ship with a pinned Qwen3-VL-8B-Instruct Q8_0 text encoder (8.71 GB) and the BF16 VAE (0.68 GB). Two non-obvious requirements: --vae-tiling is mandatory at 1024² or the SYCL VAE overflows int32, and one sd-cli per GPU — a second concurrent run wedges the xe driver and needs a root-only reset. The 3.84× speedup comes from --eager-load --cache-mode easycache, not from fewer steps.

FlagOS ships eight builds for Chinese NPUs, same 28-file layout in each:

Name	Target	Precision	Size	Links
BF16 nvidia	NVIDIA (FlagOS path)		33.13 GB
BF16 hygon	Hygon DCU		33.13 GB
BF16 ascend	Ascend NPU		33.13 GB
BF16 metax	MetaX CGC		33.13 GB
BF16 enflame	Enflame GCU		33.13 GB
BF16 zhenwu	Zhenwu MUSA		33.13 GB
BF16 mthreads	Moore Threads MUSA		33.13 GB
W8A8 arm	ARM		29.38 GB

· · · · · · · · · · · · · ·

Same weights, re-laid-out for a ComfyUI-side loader. Each ships a YAML manifest next to the shards, so no diffusers config is needed.

Name	Format	Precision	Size	Links	Notes
libwaifu bf16	libwaifu (yaml + 8 shards)		30.03 GB		Full-precision layout for the libwaifu loader.
libwaifu fp8	libwaifu (yaml + 4 shards)		15.98 GB		Same layout, weight_format: fp8. Half the download.

· · · · · · · · · · · · · ·

Two drop-in decoder swaps and one preview-only toy. The first is a real finetune; the second is not a VAE at all.

Name	Precision	Size	Links	Notes
Texture-fix VAE		0.68 GB		Drop-in replacement for qwen_image_2.1_vae_bf16.safetensors. Decoder-only finetune — 7.5M trainable params, the top two decoder stages plus the output head, ~5,000 steps at 3e-5 with a TAESD recipe (MSE + LPIPS + GAN). Targets checkerboard artifacts and flat texture in photo-style images. rFID 3.37 → 2.08 on COCO at 256², at the cost of PSNR (33.30 → 32.86) and LPIPS (0.0357 → 0.0373): sharper and more plausible, slightly less faithful. Ships a diffusers folder too.
TAE preview decoder		0.003 GB		Not a VAE — a 1.63M-param 16× approximate decoder for ComfyUI live previews, ~15 ms for 1024² on an RTX 5080. 30.1 dB against Latent2RGB's 20.9. Goes in models/vae_approx/ and needs ComfyUI-KJNodes; the built-in TAESD option will not load it, since that path only handles 8× decoders. RGB only — it drops the alpha channel, so never use it for the final decode of an RGBA generation.
hdr vae test		0.68 GB		Untested in the wild; treat as an experiment, not a drop-in replacement.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

Name	Type	Links	Notes
training assistant v1			A working pytorch_lora_weights.safetensors + training_details.json you can inspect to learn a real SimpleTuner config.
qwen_image2.1_colab			7-cell notebook with form UI and enable_model_cpu_offload(). Works on L4 22 GB; T4 16 GB will OOM by design.
qwen_image2.1_molab			Script version of the above, tuned for Marimo and Blackwell.
Qwen-Image-2.1-Skills			Turns a short scene into a structured prompt for believable casual phone photography. 22 example images.
qwen-image-2.1-p150			Tenstorrent Blackhole p150a. tt-model pull --with-weights then tt-model serve.
ComfyUI-AlphaTrace			Diagnostic sampler nodes for the native RGBA output: they emit the predicted image, the alpha map and per-step alpha statistics at every step. Purely observational. For 2.1 pass LTXVScheduler as sigmas (max_shift 0.69, base_shift 0.54, stretch on, terminal 0.02) to match the official schedule. MIT, no extra dependencies, example workflow included.
Qwen-Image-2.1 VRAM calculator			Pick the DiT, text encoder and VAE files (BF16, FP8, INT8, GGUF), where the encoder runs and the image size; shows the peak VRAM range and which 8–32 GB GPUs fit, next to peaks people measured in ComfyUI and diffusers.
Image Studio			Local web app for Apple Silicon (64 GB+): text-to-image, edits with up to 10 reference images, transparent PNGs, background removal, and a node canvas that chains steps. Runs the Uncensored Q8_0 GGUF on ComfyUI, with bf16 copies of the int8 text encoders because MPS has no int8 matmul. AGPL.

Gotchas that will save you an afternoon

1. QwenImage21Pipeline requires diffusers main — release 0.40.0 does not have it. Install from git.
2. The CFG kwarg is true_cfg_scale, not guidance_scale, and it needs a negative_prompt to engage.
3. Width and height must be multiples of 32, max edge 2048.
4. With CPU offload, the generator should live on 'cpu'.
5. The default sample is 40 steps with no CFG. Adding CFG changes the look; don't assume more steps help.
6. Chinese prompts work natively — the PE models are what rewrite them into English.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

This list is a set of Markdown files, not one. README.md is generated and should never be edited directly — your change will be overwritten on the next build. Each section lives in its own file under blocks/:

The file prefix controls the order — the build concatenates blocks/*.md in filename order, so 10_ always lands before 20_. A new section just needs a new number.