[bug]: Unable to determine model type for Flux LyCORIS models

[mld2443]

Is there an existing issue for this problem?

• I have searched the existing issues

Operating system

Linux

GPU vendor

Nvidia (CUDA)

GPU model

rtx 3080

GPU VRAM

10

Version number

5.6.2

Browser

Firefox 135

Python dependencies

No response

What happened

Invoke gives a "InvalidModelConfigException: Unable to determine model type" Error when trying to add a LyCORIS model for Flux.

What you expected to happen

I expected it to add the model.

How to reproduce the problem

Try to add a Flux LyCORIS model. Here are a few that did not work for me:
https://civitai.com/models/674729/1930s-rubber-hose-cartoon-style-cuphead-style-flux
https://civitai.com/models/684044/90s-sitcom-style-flux
https://civitai.com/models/708309/ukiyoflow

Additional context

Strangely, here is the only one I've found that has worked for me:
https://civitai.com/models/264261/rollercoaster-tycoon-styled-images

I suspect these models are trained differently.

Discord username

No response

[vrubzov1957]

I had similar problem but for Lokr models, when used NF4 Flux base model (but when try invoke generation, not when addinf). May be source of problem same (or may be not) #6998

And also when install with some Lora models I have same problem, examples lora with problems:

Error: Unable determine model type

• https://civitai.com/models/968982/zphyr-flux-immersive-photorealistic-lora

Error: Unknown LoRA type

• https://civitai.com/models/580857/realistic-skin-texture-style-xl-detailed-skin-sd15-flux1d
• https://civitai.com/models/758852/realism-cinematic-photographic-style-f1d

[mld2443]

I looked for LoKr in the open issues, looks like I missed it as it is closed and the fix appears to have been merged in already.

Just to make sure, I grabbed the 5.7.0rc1 before trying again.

Perhaps there is more work to do, as indeed some of the models that give me the "InvalidModelConfigException" mention being LoKr, such as these:
https://civitai.com/models/683222/spirited-away-ghibli-style-for-flux
https://civitai.com/models/705358/cats-who-smoke

Meanwhile, this LoHa works fine for me:
https://civitai.com/models/50940?modelVersionId=826402

[mld2443]

I've parsed the metadata, and for each LyCORIS I've found that does not work, it seems the keys are all of the form:

lycoris_single_transformer_blocks_XXXXX.alpha
lycoris_single_transformer_blocks_XXXXX.lokr_w1
lycoris_single_transformer_blocks_XXXXX.lokr_w2
...
lycoris_transformer_blocks_XXXXX.alpha
lycoris_transformer_blocks_XXXXX.lokr_w1
lycoris_transformer_blocks_XXXXX.lokr_w2

I thought to compare them to the keys inside one of my SDXL LoKr, but they seem cut from a different cloth entirely. I have a hunch that those models might be generated on tensor.art and reuploaded (except the one with smoking cats which links to huggingface and has more details on the training).