Distilling Krea2 into Anima with a LoRA trained on just 1,200 captions

Honest_Concert_6473 · reddit · 2026-10-31

• The author trained a LoRA teaching open-source image model Anima to imitate how Krea2 Turbo generates images — a lightweight form of knowledge distillation.
• Two behaviors were trained: Anima watches Krea2's full generation process (learning construction, not just outputs), and Anima drafts while Krea2 refines from an intermediate latent via Hires Fix — like an instructor correcting a student's sketch.
• The scale was surprisingly small: 1,200 captions (600 photo-derived, 600 anime-derived) sufficed for distillation plus CFG 1/12-step distillation, well within individual reach.
• The main LoRA applies to Anima Base v01; the backup ZIP includes training code and a RunPod API-automated workflow, with a non-distilled variant for CFG 4.5/30-step use.
• Caveat: Anima doesn't match Krea2's photorealism; gains are mainly in overall visual feel. Code and LoRA are on Civitai.

Original post →

More from Multimodal

• Iridescent generative art: caustic blobs, wireframe waves and flowing forms — thespite · 2026-10-31
• Podcast: solving creator attribution and taste in the age of AI — hsu_byron · 2026-10-31
• fal gives away free Flux 3 Max generations until November 2 — noahsolomon · 2026-10-31
• Genie 4 generates a guy saying "ThursdAI" in real time; Gemini 4 Carbon tops the new index — altryne · 2026-10-31
• Open-source AI image gen stuns with hyper-realistic nine-tailed fox group scene — Some-Dark-5802 · 2026-10-31
• TONGYI SpeechAI: Streaming TTS, not LLMs, is the real bottleneck for voice assistants — rdesh26 · 2026-10-31

Multimodal channel →