Bad Output
Hi, thanks for releasing this LoRA. The model card lists black-forest-labs/FLUX.2-klein-9B as the base model, but the safetensors file has key names like diffusion_model.double_blocks.0.img_attn.qkv.weight, which is Flux.1-style architecture (double_blocks). Loading it against an actual Flux.2-Klein-9B checkpoint in ComfyUI throws shape mismatch errors, e.g.:
ERROR lora diffusion_model.double_blocks.0.img_attn.qkv.weight shape '[12288, 4096]' is invalid for input of size 28311552
and a long list of lora key not loaded: lora_unet_single_blocks_*_modulation_lin... warnings — the single_blocks keys aren't matching at all.
Could you confirm whether this was actually trained on Flux.2-Klein-9B, or whether the base_model tag is incorrect and it's actually Flux.1? Would love to use this for albedo extraction but it's currently unusable against the real Flux.2 checkpoint. Happy to share the full log if useful.
Same. I tried using it on Flux-2-Klein-Base-9b on a game texture and the result doesn't represent the original texture at all. I'm going to try again on regular Flux-2-Klein-9b, but since that's what you tried I doubt it will work then either. Given the author's example images, maybe it wasn't trained on actual game textures like what we're trying.
Sure enough, also get this in ComfyUI output log:
[WARNING] lora key not loaded: transformer.single_transformer_blocks.20.attn.to_out.lora_A.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.20.attn.to_out.lora_B.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.20.attn.to_qkv_mlp_proj.lora_A.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.20.attn.to_qkv_mlp_proj.lora_B.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.21.attn.to_out.lora_A.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.21.attn.to_out.lora_B.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.21.attn.to_qkv_mlp_proj.lora_A.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.21.attn.to_qkv_mlp_proj.lora_B.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.22.attn.to_out.lora_A.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.22.attn.to_out.lora_B.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.22.attn.to_qkv_mlp_proj.lora_A.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.22.attn.to_qkv_mlp_proj.lora_B.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.23.attn.to_out.lora_A.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.23.attn.to_out.lora_B.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.23.attn.to_qkv_mlp_proj.lora_A.weight
[WARNING] lora key not loaded: transformer.single_transformer_blocks.23.attn.to_qkv_mlp_proj.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.5.attn.to_k.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.5.attn.to_k.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.5.attn.to_out.0.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.5.attn.to_out.0.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.5.attn.to_q.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.5.attn.to_q.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.5.attn.to_v.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.5.attn.to_v.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.6.attn.to_k.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.6.attn.to_k.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.6.attn.to_out.0.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.6.attn.to_out.0.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.6.attn.to_q.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.6.attn.to_q.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.6.attn.to_v.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.6.attn.to_v.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.7.attn.to_k.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.7.attn.to_k.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.7.attn.to_out.0.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.7.attn.to_out.0.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.7.attn.to_q.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.7.attn.to_q.lora_B.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.7.attn.to_v.lora_A.weight
[WARNING] lora key not loaded: transformer.transformer_blocks.7.attn.to_v.lora_B.weight
Can try this one I made specifically for game textures
https://huggingface.co/spaces/NightRaven109/PBRFusion4AlbedoKlein
Can also try out my open source pbrfusion project for normal and displacement generation
https://github.com/Night1099/COMFYUI-PBRFusion4
And/or shameless self promo this company I co-founded for a more streamlined material generation pipeline
pbrgen.com
Shameless indeed, but with good results, so thank you. For albedo generation via your first link I found that your Reinhard color matching generates an incorrect albedo: I'd default to leaving it disabled, which looks much better.
It's the strength of the model; I did the test and it gives me the same result. If you lower its strength, for example to 0.25/0.30, it gives results similar to the samples.
It's the strength of the model; I did the test and it gives me the same result. If you lower its strength, for example to 0.25/0.30, it gives results similar to the samples.
Hi, I'm experiencing the same issues with comfyui. I lowered the strength and got better results. I'm going to share the python script I use, can anyone try with it?
Same. I tried using it on Flux-2-Klein-Base-9b on a game texture and the result doesn't represent the original texture at all. I'm going to try again on regular Flux-2-Klein-9b, but since that's what you tried I doubt it will work then either. Given the author's example images, maybe it wasn't trained on actual game textures like what we're trying.
If you give me some textures of yours, i can do inference and return to you.
No need, tried it at the lower strength and now it does indeed work. Nice lora, it works great.
Hi, I'm sharing the source code for inference in gradio in the home page.