No trigger word is needed for this LoRA. Though using something like <code>This is a absurdres, high quality, aesthetic masterpiece cinematic anime artwork.</code> or something around those lines helps Chroma produce better images.I invite you to add other LoRA to make the image quality better. The showcase images were generated with showcasing the artist style, not the wacky combination you can use it in (comfy workflow is included in the images). You can easily get better looking result if you try a bit harder, this is only meant to showcase the effect of the LoRA alone.Captioning was done <a target="_blank" rel="ugc" href="https://huggingface.co/mradermacher/Qwen3-VL-8B-NSFW-Caption-V4.5-GGUF">Qwen3-VL-8B-NSFW-Caption-V4.5</a> which is better than joycaption at everything but hardcore content.<hr />v2.1 looks almost identical to v2 (it might be 5% worse in therm of pure styling), but is way less over-fitted and was trained with better settings which should help with diversity. It is also way more stable, you can push the LoRA strength to 1.5 and still get decent results.

Style > PossumMachine

Lora trained on 60 PossumMachine images at 1MP, using kohya-ss/sd-scripts with Adamw -1e04 LR constant for ~5k stepsThere is a lot of room for improvement by changing the LR to something smaller at the end, adding a reg dataset, masking image with white background, ...The captioning is also mediocre at best (JoyCaption makes a LOT of mistakes).

Files with SHA256: 1a86e8c1beab5da7bc2fd1c92d3acd32cf747e4949e789d5e5730339b1bcdc50