tutorials/image/qwen/qwen-image-layered

[giscus[bot]]

tutorials/image/qwen/qwen-image-layered

Qwen-Image-Layered is a model capable of decomposing an image into multiple RGBA layers, enabling inherent editability through layer decomposition.

https://docs.comfy.org/tutorials/image/qwen/qwen-image-layered

[MatzGen]

Something seems to be odd, ksampler freezes at 78% and in the subgraph only the seed setting which is set to randomize by default got a violet border.

I updated everything (comfy and all nodes) and it doesn't appear to be anything missing. Any hints maybe?

Console doesn't show any errors... using the bf16 on a RTX 4090 24 GB

got prompt
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
Requested to load WanVAE
loaded completely; 21736.67 MB usable, 242.04 MB loaded, full load: True
Found quantization metadata version 1
Using MixedPrecisionOps for text encoder
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load QwenImageTEModel_
loaded completely; 21434.64 MB usable, 7910.29 MB loaded, full load: True
model weight dtype torch.bfloat16, manual cast: None
model_type FLUX
Requested to load QwenImage

[MatzGen]

I was wondering because it says WanVAE and model_type FLUX in the console, so I checked load diffusion model is default weight dtype and load clip type is definitely qwen image, also the vae is set to qwen layered vae, so not sure if this is normal / where this comes from.

[MatzGen]

using the fp8 model it runs through till the vae decode and then crashes WHEN weight_dtype is set to "default".

When weight_dtype is set to fp8_e4m3fn it runs through and gives an output.

Cheers.