Better Build: This version uses smarter tweaks for keeping images looking consistent and stable, especially on tricky edits, beating the original Qwen-Image-Edit.
Nicer Looks: Adds tools and data to amp up colors, contrast, lighting, and art vibe, making pics more natural and pro-level.
Keeps Details Sharp: Holds onto textures, faces, and text better, great for all kinds of pics like people, scenes, products, or drawings, whether changing meaning or just style.
Try these prompts to explore the model’s strengths:
“make the lighting soft and cinematic with better balance”
“enhance the photo’s composition and maintain realism”
“refine skin tone and texture consistency”
“improve the global color tone and aesthetic harmony”
“increase photo realism and clarity without changing content”
These models are redistributed here for the sake of convenience.
Description
FAQ
Comments (23)
How does this compare to just using the lora?
Qwen-Image-Edit 20 billion parameters vs. 3 million trainable LoRA parameters.
@sweetmax797 the 3 million trainable lora parameters also uses the 19.997 billion other parameters, otherwise we could just run the 3 million on its own without Qwen at all. I'm asking about outcome. Numbers don't communicate how it compares visually.
@Jellai It means it is a checkpoint and can't be compared to a LoRA. I'm still testing it, but I can say it's a bit more precise compared to regular FP8. However, compared to BF16, aside from visual effects and filters, they're the same. (but to early to draw a conclusion)
@sweetmax797 Okay, but many people compare checkpoints to loras, when the lora is extracted from the checkpoint, and we use loras extracted from checkpoints all the time. A few days ago, someone uploaded a lora that was extracted from this. Then you uploaded the whole checkpoint, so I wanted to know how much of a difference there was, if it was worth downloading 19GB instead of using the lora.
I'm very certain what I'm asking is not silly. I'm not confused about the difference between loras and checkpoints.
The question is more how is compared to original Qwen Image Edit 2059 model.
@Jellai Sorry, I didn't mean to play teacher, so your question is by no means silly. I'll check the metadata, and if there's a sign of a merge, I'll share the retrieved information. But according to the model info, it's a fine-tuned model
@sweetmax797 Oh it's okay. I'm certain this is a fine tuned model too. I guess if no one has checked the difference, then the only thing left is for me to try it myself. Thanks.
Oh, and here's the extracted lora I was talking about:
https://civitai.com/models/2075957/qweneditmeitulorar32
@Jellai Just checked the metadata, there are none. The claim that you can extract LoRA from it is simply false. There are no text encoder keys like clip_1 or t5, VAE keys, or LoRA keys (like lora_up or lora_down). The only way to extract LoRA from a merged model is if it’s merged using the concat method if that method is used, the model info on Hugging Face would show it. If LoRA is merged via the linear method, it is impossible to even tell if it’s merged. However, discussion about extracting LoRA is completly false.
this article and codes can clarify it more : https://civitai.com/articles/18798/qwen-image-nsfw-lora-notes
metadata wasn't present but these keys can tell a lot.
--- Tensor Keys (Model Layers/Parameters) ---
Found 1933 tensors: <- flux DiT
- model.diffusion_model.img_in.bias
- model.diffusion_model.img_in.weight
- model.diffusion_model.norm_out.linear.bias
- model.diffusion_model.norm_out.linear.weight
- model.diffusion_model.proj_out.bias
- model.diffusion_model.proj_out.weight
- model.diffusion_model.time_text_embed.timestep_embedder.linear_1.bias
- model.diffusion_model.time_text_embed.timestep_embedder.linear_1.weight
- model.diffusion_model.time_text_embed.timestep_embedder.linear_2.bias
- model.diffusion_model.time_text_embed.timestep_embedder.linear_2.weight
...
- model.diffusion_model.transformer_blocks.9.img_mod.1.weight
- model.diffusion_model.transformer_blocks.9.txt_mlp.net.0.proj.bias
- model.diffusion_model.transformer_blocks.9.txt_mlp.net.0.proj.weight
- model.diffusion_model.transformer_blocks.9.txt_mlp.net.2.bias
- model.diffusion_model.transformer_blocks.9.txt_mlp.net.2.weight
- model.diffusion_model.transformer_blocks.9.txt_mod.1.bias
- model.diffusion_model.transformer_blocks.9.txt_mod.1.weight
- model.diffusion_model.txt_in.bias
- model.diffusion_model.txt_in.weight
- model.diffusion_model.txt_norm.weight
@Jellai And I checked the LoRA; thanks for the link, by the way. I found an interesting piece of information: the LoRA has the key - diffusion_model.img_in.diff_b.
This means you need the base model plus the fine-tuned model to extract the differences, but this isn't an ordinary task and requires expertise. Therefore, the model is not a merged model, this LoRA is essentially stolen work. That is the only way that key would ever be there. However, since it's open source, at least he could tell the truth and appreciate the work and fine-tuning.
/ComfyUI/models/loras/ComfyUI_meitu_lora_rank_32_fp16_00001_.safetensors ---
No metadata found in this file.
--- Tensor Keys (Model Layers/Parameters) ---
Found 2779 tensors:
- diffusion_model.img_in.diff_b
- diffusion_model.img_in.lora_down.weight
- diffusion_model.img_in.lora_up.weight
- diffusion_model.norm_out.linear.diff_b
- diffusion_model.norm_out.linear.lora_down.weight
- diffusion_model.norm_out.linear.lora_up.weight
- diffusion_model.proj_out.diff_b
- diffusion_model.proj_out.lora_down.weight
- diffusion_model.proj_out.lora_up.weight
- diffusion_model.time_text_embed.timestep_embedder.linear_1.diff_b
...
- diffusion_model.transformer_blocks.9.txt_mlp.net.2.diff_b
- diffusion_model.transformer_blocks.9.txt_mlp.net.2.lora_down.weight
- diffusion_model.transformer_blocks.9.txt_mlp.net.2.lora_up.weight
- diffusion_model.transformer_blocks.9.txt_mod.1.diff_b
- diffusion_model.transformer_blocks.9.txt_mod.1.lora_down.weight
- diffusion_model.transformer_blocks.9.txt_mod.1.lora_up.weight
- diffusion_model.txt_in.diff_b
- diffusion_model.txt_in.lora_down.weight
- diffusion_model.txt_in.lora_up.weight
- diffusion_model.txt_norm.diff
@sweetmax797 is it an advantage downloading~using meitu model instead of qie2509 original or it's not ?
@amazingbeauty I haven't really tested it out. It depends on whether the method they used to extract the differences between the base model and MeiTu is robust or not. The only way to find out is to actually test them. If the results are identical and you've already downloaded the base model, the advantage is that you save some disk space.
@sweetmax797 your full model or qie2509 model is better at realistic photo editing?
@amazingbeauty realistic? If you mean a 'Fluxy' or 'plastic' look, it relates to the type of sampler used. Euler is fast but gives a plastic render since it's primarily for testing. If you want realistic skin, you need to use the 'DPM++' family with 'Karras.'". and about the LoRA, I did one render yesterday after we talked. It was good, but needs more testing scenario, so I'll do several scenario test and share the result.
Really need a gguf copy for my 12 GB Vram
12GB vram + 64GB Rram, u can run bf16 up to 40GB model size, it's more than enough for fp8. the GGUF download link is in the description, first line
@sweetmax797 running in RAM is extremely slow, only when models fits in VRAM its possible to use without extreme waiting.
I run this model - no gguf ever - on a 3060 12GB vram and 32GB ram.
this is a typical finetune trained on a dataset? Or it's some DIT aesthetic guided sorcery?
What actually is it, and is it an all rounder that lets me delete my typical 2509 edit? Because I don't have enough drive space for it otherwise.
There is a link in Jellia's comment pointing to a LoRA that extracted the differences between the base model and this one. I haven't tried it yet, so you could try that LoRA first. For me, deleting and re-downloading goes faster than archiving them on a USB or HDD. So far, I’ve used it with faces and realism, it's not as effective with illustrations and some digital art. info links are in the description if you wanna dig more.
Friend, could you please provide a workflow? The workflow for this version of the model, and the workflow for the gguf version model? Thank you so much, and may you have a peaceful and prosperous life.
Thank you for the kind words and you too. the ComfyUI default Qwen-Image-Edit works with it. all you need to do is just select this model in diffusion loader node. but if you look at sidebar on the right where image info is, you can copy the workflow I use ( https://imgur.com/rPtChHC ) just click on copy icon and "CTRL + V" in ComfyUI canvas. But i use some custom nodes that is coded by me for diffusion loader (Qwen-Image Advanced Diffusion Loader ) and it isn't shared on custom node repository. Replace it with the ComfyUI's core diffusion loader. the rest of the nodes like AIO Aux Preprocessor or Qwen-Image Size Picker, can be installed via custom node manager.
So its based on old Image-Edit, not on 2509?