Image to Image - Qwen i2i

#58
by leolis78 - opened

Hi, I’ve been testing img2img (image-to-image) with Qwen-Image in ComfyUI and found a consistent issue:
When starting from a pure noise latent, the generation works as expected: detailed, sharp, and visually consistent.
When starting from an encoded latent (VAE Encode of an input image), the output often looks washed out / watercolor-like, and in many cases exhibits visual artifacts or partially broken images, even at higher denoise values.
This does not happen when I use the same workflow with other models (e.g., Flux), where img2img from an encoded latent preserves or improves detail without breaking.

Steps to Reproduce:
Load Qwen-Image FP8/BF16 (also tested with distilled).
Use official Qwen VAE for both encoding and decoding.
Encode a real image → VAE Encode → KSampler (img2img).

Settings tested:
Sampler: Euler / Res_multistep
Scheduler: Simple / Beta
Steps: 28–36 (FP8/BF16), 10–15 (distilled)
CFG: 1–6 (FP8/BF16), 1.0 (distilled)
Denoise: 0.25–0.6
ModelSamplingAuraFlux shift: 1–3.1

Compare outputs:
Case A: latent = pure noise (same seed) → sharp, detailed, artifact-free. :)
Case B: latent = VAE Encode of an image → watercolor-like texture and sometimes broken visuals (patchy areas, corrupted colors, loss of structure) :(

Observed Behavior:
From encoded latents:
Loss of high-frequency detail (watercolor/plastic look)
Structural artifacts (patchy regions, color corruption, distorted geometry)

Is there a known limitation or workaround for Qwen-Image when doing img2img from encoded latents?
Can you help me? Thanks in advance!

Hi, can you show your workflow?

This comment has been hidden

Hi, can you show your workflow?

Hi etaooo, I attach the json

JSON = https://pastebin.com/f5WCsPdx

ComfyUI_01083_.png

image.png

I also tested the MaskPaint node and it shows the same problem.
For me, MaskPaint is very important because I use it to add fine details in specific areas such as the face, eyes, hands, etc.

Regards

Hi, can you show your workflow?

Hi etaooo, I attach the json

JSON = https://pastebin.com/f5WCsPdx

ComfyUI_01083_.png

image.png

I also tested the MaskPaint node and it shows the same problem.
For me, MaskPaint is very important because I use it to add fine details in specific areas such as the face, eyes, hands, etc.

Regards

I did some testing with fp8 and here is the result, ( Iuse dpmpp_2s_ancestral and ddim_uniform [Normal is better for text tho] )

input img:

01.jpg

Output

ComfyUI_temp_toayf_00002_.png

Sign up or log in to comment