Roadmap

[Jasonzzt]

ComfyUI-CacheDiT TODO List

• Fix model restore on disabling node feat: add enable/disable functionality for cache optimizers #28
• Extend model support e.g. Flux.2 feat: Add Flux series support and improve cache handling in models #33
• Canceling batch tasks error
• Fix last step computation

[NRTnarathip]

could you add support for Flux2 Kelin 4b, 9b .gguf & .safetensors

[Jasonzzt]

could you add support for Flux2 Kelin 4b, 9b .gguf & .safetensors

Sure, next work is to support Flux2 Kelin 4b, 9b

[Moussack]

could you add support for Flux2 Kelin 4b, 9b .gguf & .safetensors

Sure, next work is to support Flux2 Kelin 4b, 9b

idk, isn't the current version is already worked with klein 9B?, i have 50-60% speedup using klein 9B distilled with cacheDIT.
i'm using the zimage turbo setup.

[hum-ma]

How about Anima? There is an error "BlockAdapter has no transformer reference" but it still processes the workflow, and also shows "BlockAdapter verified: 28 blocks" and skips the correct steps. I'm not certain if the result is as it should be though, quality doesn't really seem better than simply lowering total number of steps by the same amount.

Another thing, shouldn't the last skip generally be overridden and always computed in the case that it falls on the very last step? Currently it may end up skipping from e.g. 18/20 to 20/20 done, and the resulting image will have leftover noise.

[Jasonzzt]

How about Anima? There is an error "BlockAdapter has no transformer reference" but it still processes the workflow, and also shows "BlockAdapter verified: 28 blocks" and skips the correct steps. I'm not certain if the result is as it should be though, quality doesn't really seem better than simply lowering total number of steps by the same amount.

Another thing, shouldn't the last skip generally be overridden and always computed in the case that it falls on the very last step? Currently it may end up skipping from e.g. 18/20 to 20/20 done, and the resulting image will have leftover noise.

Thanks for the feedback.

Regarding Anima support: Currently, Anima is not officially supported. The error BlockAdapter has no transformer reference confirms that the current implementation cannot correctly hook into Anima's specific Transformer architecture. While the logs show it's "skipping" steps, it's not effectively caching the internal features as intended, which is why you aren't seeing quality improvements. I don't have plans to support Anima at this time.

Regarding the "Last Step" skip: You've made an excellent point. In my current implementation, if the skip_interval happens to align with the final step, it will indeed reuse the previous latent instead of performing the final denoising. This can leave residual noise that affects VAE decoding.

I will update the logic to force a full computation on the very last step, regardless of the skip interval. This should ensure the final image is fully converged and clean.

[NRTnarathip]

is will support Flux2 Klein Distilled? or just base version?

[Jasonzzt]

is will support Flux2 Klein Distilled? or just base version?

Only base model :)