Qwen-Image-Edit 2511 Turbo — GGUF build
20B instruction-driven image editing with the Lightning 8-step LoRA pre-merged. CFG-free; the transformer fits a single 24 GB card.
What this repo is: a requantized, split-layout build of the upstream model — weights only, not a retrain. The recommended settings below are the values these weights are actually run with day to day, not copied from the upstream card.
Samples
img2img
This model does img2img as well as txt2img. The left image is the input; the right is what this model made of it.
No strength is set, and none is needed. This is an instruction-edit model: it reads the prompt as an edit to apply rather than denoising the source by a fraction. A sweep at 0.3 / 0.5 / 0.7 / 0.9 produced four near-identical outputs — strength is inert here, which is why the preset omits it.
That is also why this model preserves the scene most faithfully of the family: the window, the table, the apple's position and its cast shadow all stay where they were while the surface becomes thick impasto brushwork. The others trade composition for style as strength rises; this one does not have to.
Recommended settings
Values this build is run with in practice. They are read out of a live config when this card is generated, so they cannot drift from what is actually used.
| Parameter | Production value | Meaning |
|---|---|---|
steps | 8 | Denoising steps |
guidance | 1.0 | Guidance scale |
true_cfg_scale | 1.0 | True-CFG scale (>1 enables the negative branch) |
Supported modes: txt2img, img2img
Notes and gotchas
true_cfg_scaleis pinned at 1.0. The Lightning distillation is CFG-free: at 1.0 the negative branch is disabled and each step is one transformer forward, versus two above 1.0. That makes 1.0 unambiguously right for speed. Whether raising it would improve edit quality has not been measured on this build.- Raising
stepsdoes not help. 8, 12 and 16 were rendered from the same source image, instruction and seed and are visually indistinguishable — the extra steps only cost time (~81s / ~122s / ~162s per edit on an A100). 8 is what the Lightning LoRA is distilled for. - Edit texture loss was a quantisation problem, and Q6_K largely fixes it. On the previous Q4_K_M build, edits returned the subject with its surface detail smoothed and colours re-saturated — a hat edit turned the source apple's speckled, droplet-covered skin into flat glossy red. Step count was ruled out first (8, 12 and 16 are visually indistinguishable), leaving quantisation. Re-rendering every sample on Q6_K at the identical preset restored visible speckle: edge variance rose from 103.7 to 126.8 on the hat edit and 693.1 to 895.1 on the background swap, and the samples above are the Q6_K renders. Caveat on how tight that measurement is: the Q4 edits were made from an earlier source image that no longer exists, so the two sets were not driven from a byte-identical input — the effect is large and in the predicted direction, but it is not a perfectly controlled A/B.
guidanceis inert on this build (guidance_embeds=falseon the transformer config) but is kept explicit for clarity.- No
width/heightin the preset — this is deliberate. Omitting them lets the pipeline preserve the input image's aspect ratio. - There is no
strengthparameter — this build has no partial-denoise branch. - Fallback if you need the base build: transformer
qwen-image-edit-2511-Q4_K_M.ggufwith{steps: 40, guidance: 1.0, true_cfg_scale: 4.0}.
Layout
Components ship as separate files rather than one bundle: the transformer, the text encoder(s) and the VAE each load independently, with configs and tokenizers under config/. Any loader that accepts explicit per-component paths can consume this directly — point it at the files listed below.
A generic
diffuserssnippet is deliberately omitted. This is a repacked split layout with substituted encoders, so an upstream example will not load it unmodified, and an untested snippet would be worse than none.
Files
| File | Size | Role |
|---|---|---|
split/text_encoders/Qwen2.5-VL-7B-Instruct-q4_0.gguf | 4.13 GB | text encoder |
split/text_encoders/Qwen2.5-VL-7B-Instruct-mmproj-f16.gguf | 1.26 GB | vision projector (mmproj) |
split/qwen_image_vae.safetensors | 242.05 MB | VAE |
config/processor/tokenizer.json | 10.89 MB | tokenizer / processor |
config/tokenizer/tokenizer.json | 6.71 MB | tokenizer / processor |
config/tokenizer/vocab.json | 3.23 MB | tokenizer / processor |
config/processor/vocab.json | 2.65 MB | tokenizer / processor |
config/tokenizer/merges.txt | 1.59 MB | tokenizer / processor |
config/transformer/diffusion_pytorch_model.safetensors.index.json | 194.23 KB | transformer (main weights) |
config/text_encoder/model.safetensors.index.json | 56.30 KB | text encoder |
config/processor/tokenizer_config.json | 4.62 KB | tokenizer / processor |
config/tokenizer/tokenizer_config.json | 4.58 KB | tokenizer / processor |
config/text_encoder/config.json | 3.07 KB | text encoder |
config/tokenizer/config.json | 1.34 KB | tokenizer / processor |
config/processor/video_preprocessor_config.json | 910 B | tokenizer / processor |
config/processor/preprocessor_config.json | 826 B | tokenizer / processor |
config/vae/config.json | 730 B | VAE |
config/processor/special_tokens_map.json | 613 B | tokenizer / processor |
config/tokenizer/special_tokens_map.json | 613 B | tokenizer / processor |
config/processor/added_tokens.json | 605 B | tokenizer / processor |
config/tokenizer/added_tokens.json | 605 B | tokenizer / processor |
config/model_index.json | 516 B | config |
config/scheduler/scheduler_config.json | 485 B | scheduler config |
config/transformer/config.json | 362 B | transformer (main weights) |
config/tokenizer/preprocessor_config.json | 350 B | tokenizer / processor |
config/text_encoder/generation_config.json | 244 B | text encoder |
Provenance
- Upstream base model:
Qwen/Qwen-Image-Edit-2511(upstream license: apache-2.0) - This build: requantized and relaid out into a split component layout. Weights are not retrained; only the format and directory layout differ.
- License: left as
unknownin this repo's metadata. Refer to the upstream model's license for redistribution and commercial-use terms — several of these bases are non-commercial.