[enhancement]: Support for Flux.2 model

[sammcj]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

No response

What should this feature add?

It would be great if support for Flux.2 could be added to Invoke.

https://huggingface.co/black-forest-labs/FLUX.2-dev

Alternatives

No response

Additional Content

No response

[sammcj]

I see that Flux.2 Klein (smaller, fast model) now works 🎉, but FLUX.2 Dev (larger, higher quality model) is still marked as unidentified and when you edit to to be a main model the UI errors (but doesn't say what the error is) and the console shows a pydantic validation error:

Input tag 'main.unknown.unknown' found using get_model_discriminator_value() does not match any of the expected tags: 'main.diffusers.sd-1', 'main.diffusers.sd-2', 'main.diffusers.sdxl', 'main.diffusers.sdxl-refiner', 'main.diffusers.sd-3', 'main.diffusers.flux', 'main.diffusers.flux2', 'main.diffusers.cogview4', 'main.diffusers.z-image', 'main.checkpoint.sd-1', 'main.checkpoint.sd-2', 'main.checkpoint.sdxl', 'main.checkpoint.sdxl-refiner', 'main.checkpoint.flux2', 'main.checkpoint.flux', 'main.checkpoint.z-image', 'main.bnb_quantized_nf4b.flux', 'main.gguf_quantized.flux2', 'main.gguf_quantized.flux', 'main.gguf_quantized.z-image', 'vae.checkpoint.sd-1', 'vae.checkpoint.sd-2', 'vae.checkpoint.sdxl', 'vae.checkpoint.flux', 'vae.checkpoint.flux2', 'vae.diffusers.sd-1', 'vae.diffusers.sdxl', 'vae.diffusers.flux2', 'controlnet.checkpoint.sd-1', 'controlnet.checkpoint.sd-2', 'controlnet.checkpoint.sdxl', 'controlnet.checkpoint.flux', 'controlnet.checkpoint.z-image', 'controlnet.diffusers.sd-1', 'controlnet.diffusers.sd-2', 'controlnet.diffusers.sdxl', 'controlnet.diffusers.flux', 'lora.lycoris.sd-1', 'lora.lycoris.sd-2', 'lora.lycoris.sdxl', 'lora.lycoris.flux', 'lora.lycoris.z-image', 'lora.omi.sdxl', 'lora.omi.flux', 'lora.diffusers.sd-1', 'lora.diffusers.sd-2', 'lora.diffusers.sdxl', 'lora.diffusers.flux', 'lora.diffusers.z-image', 'control_lora.lycoris.flux', 't5_encoder.t5_encoder.any', 't5_encoder.bnb_quantized_int8b.any', 'qwen3_encoder.qwen3_encoder.any', 'qwen3_encoder.checkpoint.any', 'qwen3_encoder.gguf_quantized.any', 'embedding.embedding_file.sd-1', 'embedding.embedding_file.sd-2', 'embedding.embedding_file.sdxl', 'embedding.embedding_folder.sd-1', 'embedding.embedding_folder.sd-2', 'embedding.embedding_folder.sdxl', 'ip_adapter.invokeai.sd-1', 'ip_adapter.invokeai.sd-2', 'ip_adapter.invokeai.sdxl', 'ip_adapter.checkpoint.sd-1', 'ip_adapter.checkpoint.sd-2', 'ip_adapter.checkpoint.sdxl', 'ip_adapter.checkpoint.flux', 't2i_adapter.diffusers.sd-1', 't2i_adapter.diffusers.sdxl', 'spandrel_image_to_image.checkpoint.any', 'clip_embed.diffusers.any.gigantic', 'clip_embed.diffusers.any.large', 'clip_vision.diffusers.any', 'siglip.diffusers.any', 'flux_redux.checkpoint.flux', 'llava_onevision.diffusers.any', 'unknown.unknown.unknown' [type=union_tag_invalid, input_value={'key': 'e37f86a2-59cc-45...alse, 'config_path': ''}, input_type=dict]

[Pfannkuchensack]

The Flux2 Model (not Klein) uses Mistral as Text Encoder and there is not yet support for that in invoke. I did some testing but i cant get it to run on my 24GB vram card. I need to find a version that can fit into my card then i will add support for it. There is a Node for it https://discord.com/channels/1020123559063990373/1458403901941354578

[sammcj]

I wonder if we need to get it quantised to GGUF first, maybe a Q8 or even Q6_K_XL to save enough VRAM.

A slight aside from this issue, part of the beauty of GGUF is that it can run split across multiple GPUs, I don't think InvokeAI can do this yet but perhaps Invoke could place the text encoder and any adapters, LoRAs etc.. on other GPUs if detected.