Looks like someone beat z-image to the punch in respect to a dedicated Anime style model. Very lightweight too, insanity. Even on a gaddamn Sunday we're getting new releases.
https://x.com/ModelScope2022/status/1997543466587636209?t=kA4UmA71XRPsyJS515WaOQ&s=19

https://redd.it/1pgdzf1
@rStableDiffusion

NewBie Image Exp0.1: a 3.5B open-source ACG-native DiT model built for high-quality anime generation
https://modelscope.cn/models/NewBieAi-lab/NewBie-image-Exp0.1

https://redd.it/1pgehp8
@rStableDiffusion

AI-Toolkit: Use local model directories for training

For AI-toolkit trainings, I propose to download the models manually and store them locally, outside huggingface cache. This should work for all training types and usually prevents the need for online connection at the beginning of each training.

**Example for Z-Turbo with training adaptor LoRa, but the process is the same for any other training:**

1. Go to [https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/tree/main](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/tree/main) and download the folders marked in the sceenshot (text\_encoder, tokenizer, transformer, vae).
2. Store this directory structure to a dedicated training models folder, in my case "**g:\\Training\\Models\\Tongyi-MAI--Z-Image-Turbo\\**"
3. Go to [https://huggingface.co/ostris/zimage\_turbo\_training\_adapter/tree/main](https://huggingface.co/ostris/zimage_turbo_training_adapter/tree/main) and download one or both of the training adaptors zimage\_turbo\_training\_adapter\_v1.safetensors or zimage\_turbo\_training\_adapter\_v2.safetensors. After some training tests I am still not sure if V1 of or V2 works better. I tend to say V1.
4. Store the LoRas to the dedicated training models folder, in my case "**g:\\Training\\Models\\ostris--zimage\_turbo\_training\_adapter\\**"
5. Create a new job, set the correct training type and for the models enter the path to the downloaded models in this format: "**g://Training//Models//Tongyi-MAI--Z-Image-Turbo**" and "**g://Training//Models//ostris--zimage\_turbo\_training\_adapter//zimage\_turbo\_training\_adapter\_v1**"
6. Select the training dataset and make other changes as needed, then save the job.

https://preview.redd.it/f024xhmper5g1.png?width=1731&format=png&auto=webp&s=b78cd06e4e891c89deb2bb542d89dc21e91b509b

This setup also prevents the anoying re-downloads of the complete model set if minor changes happen in Huggingface repository, e.g if the readme file is updated. This results in the download of a new snapshot each time into the .cache\\huggingface\\hub\\ folder, creating duplicate data.

If you have donwloaded the models already ealier to .cache\\huggingface\\hub\\ folder via the AI-Toolkit, you can just copy/move the folders to your dedicated training models folder, and set the local paths in training setup as described above.


Finally, if you need a really comprehensive overview and explanation of latest the AI-Toolkit training settings, I can recommend this video: [https://www.youtube.com/watch?v=liFFrvIndl4&t=2s](https://www.youtube.com/watch?v=liFFrvIndl4&t=2s)
This video was done for ZImage but the detailed settings denoscriptions are relevant for all tryining types.

https://redd.it/1pgfkoa
@rStableDiffusion

Z-Image trainer that can train the distilled version of LoRA (in 4~8 steps)

https://redd.it/1pgjpec
@rStableDiffusion

DX8152's Qwen Edit 2509 Light transfer lora is out!

https://redd.it/1pgl821
@rStableDiffusion

🚀 ComfyUI_StarNodes v1.9.2 is out! ✨

Hey folks, just pushed a fresh update of StarNodes and wanted to share what’s new. 😊

https://preview.redd.it/r4yhqrzn8s5g1.png?width=2048&format=png&auto=webp&s=046c50d0b09afb352d1156b1d5d672d36b1ec217

**New nodes in 1.9.2:**

* ⭐ **Star Stop And Go** – Lets you pause your workflow, preview results, and then decide if you want to continue, pause, or bypass, so you don’t waste time on bad runs.
* ⭐ **Star Model Packer** – Combines split `.safetensors` model shards into one file and converts them to FP8 / FP16 / FP32 in a single, convenient node.
* ⭐ **Star FP8 Converter** – Takes an existing `.safetensors` checkpoint and converts it to FP8 (`float8_e4m3fn`), saving it into your standard ComfyUI output models folder for easy use.

On top of that, **a bunch of issues have been fixed** and the docs/versions are cleaned up so things should feel a bit smoother overall. 🧹✅

You can install/update **via ComfyUI Manager** (just search for “Starnodes”)
or check out the full details and docs on GitHub:
👉 [https://github.com/Starnodes2024/ComfyUI\_StarNodes](https://github.com/Starnodes2024/ComfyUI_StarNodes)

https://preview.redd.it/ge2e3lwp8s5g1.png?width=1545&format=png&auto=webp&s=258593f4990ae52dc9bbbc479178f35bf5a71307

Thanks for all the feedback and bug reports – it really helps make these nodes better for everyone. 💛

https://redd.it/1pgi6el
@rStableDiffusion

I trained Z-Image lora with prodigy-plus-schedule-free and it seems to work.

https://redd.it/1pgkxyq
@rStableDiffusion

z-image is soooo good!!!! can't wait to finetune the base

https://redd.it/1pgrwoa
@rStableDiffusion