Flux Dev8

[khadijaalfaqeer-glitch]

I'm having an issue using FLUX.1-dev8 in ComfyUI.

Whenever I run a simple txt2img or img2img workflow, I get conditioning
shape mismatch errors during sampling.

Errors:

• "size mismatch for shared.weight (256384x4096 → 32128x4096)"
• "mat1 and mat2 shapes cannot be multiplied (154x2048 and 4096x3072)"

It happens whenever any non-FLUX text encoder is used (CLIPTextEncode,
DualCLIPLoader, etc.). The model only works if all conditioning nodes
are FLUX-native, and even then the workflow is very sensitive.

Could you clarify:

1. Which text encoder nodes are officially supported for FLUX.1-dev8?
2. Is there a recommended minimal workflow?
3. Can ComfyUI show a warning when incompatible encoders are used?

Thank you!

[rehan243]

Oh, quick Q — are you sure the conditioning nodes are outputting shapes FLUX expects? That first error about shared.weight screams embedding size mismatch to me. FLUX.1-dev8 might be hardcoding its text encoder dimensions (wouldn't be surprising if it's tuned for 32128 features, not the usual CLIP defaults).

We ran into similar headaches integrating custom CLIP variants in RAG pipelines — turns out some models assume very specific dimensions and freak out if anything's off. For now, I'd double-check that you're using the exact FLUX-native encoder or see if you can shim the embeddings to match its shape.

Also yeah, ComfyUI throwing an explicit warning would definitely save time here — feels like this should be a hard validation step at graph build time instead of letting it fail during sampling.