calcuis

/

qwen-image-gguf

like 18

not sure how you config the parameters; you could just try ggc q5

even 4/8, the quality is good (see examples)
https://huggingface.co/calcuis/qwen-image-gguf/discussions/2

Well im using comfyui and i tested the non-lite normal q2 with the 4 step lightning lora and the images generated were pretty bad but i tested the same quant from city96 with the same lora i got much better results. i don't know but the issue may be pig?!

merged model might be something changed and not always follow the prompt; we can't reproduce the comfyui demo anime girl picture but got a realistic girl instead with the lite model file; that lightx lora affects the results probably

we don't understand what do you mean bad/poor quality; please at least provide some output examples with the same prompt for comparison etc.

Sorry for not including the images before. The first image generated with lite and the second image generated with city96 Q2 with lightx lora.

Here the first image with your non-lite q2 quant with lightx lora and second image is with city96 Q2 with lightx.
In the first image there's this weird grid like pattern appearing across the whole image and the quality also looks worse. This effect is more prominent in the comment above.

which quant did you pick? q2 as well?

the original model + lora should be better than the merged model; a certain degrade in the merged file as you can add up the file size is not equal

you could opt the one without distill or lite and apply a lora, then the results should be identical

non-lite q2 quant

which non-lite q2 quant? there are many q2s: q2_k, q2_k_s, iq2_s, tq2_0, etc.

non-lite q2 quant

which non-lite q2 quant? there are many q2s: q2_k, q2_k_s, iq2_s, tq2_0, etc.

It was q2_k_s.

non-lite q2 quant

which non-lite q2 quant? there are many q2s: q2_k, q2_k_s, iq2_s, tq2_0, etc.

It was q2_k_s.

of course, q2_k is better than q2_k_s; can just judge by file sizes simply, then you can estimate which one would give you a better result

which quant did you pick? q2 as well?

the original model + lora should be better than the merged model; a certain degrade in the merged file as you can add up the file size is not equal

you could opt the one without distill or lite and apply a lora, then the results should be identical

Yeah i chose q2_k_s, the one without distill or lite for the above coffee shop test.

The island image test i used the lite q2_s

very good; you could click the file to view the tensor details inside and evaluate the differences between tensor quant types; do the comparison yourself then eventually you will pick the one suits you

Okay.

Also forgot to mention the same grid like pattern did appeared when i tried to run Quantstack's distilled Q3_K_S (https://huggingface.co/QuantStack/Qwen-Image-Distill-GGUF/blob/main/Qwen_Image_Distill-Q3_K_S.gguf) with the lightx lora.

could try to fix the seed; use the same prompt to test across different quant type files; see this is a random effect or something else; not sure yet

Ok just tested both yours and city96 Q2_k with the lightx lora with same seed and prompt and the results seems similar and okay. The problem persists only when using your q2_k_s and the lite version(without lightx), especially the lite version.

in the q2_k_s attempt, only block 0 was upgraded to q4, others are all in q2 status; it might not even comparable with its q1 counterparts; that's expected.

Ok.

btw, if you try to run the lite model with gguf-connector; it do follows the prompt closely and given a good quality output as well; there are certainly a discrepancy in between