ComfyPilotComfyPilot
模型库 / LoRA / Koichame Style — Dreamy Pastel Anime
Koichame Style — Dreamy Pastel Anime

Koichame Style — Dreamy Pastel Anime

3040134

LoRA 121 次下载 11 喜欢 更新于 2026-11-21

底模:Krea 2

style

下载模型 在源站打开

触发词

@koichame

文件清单 线路:原始(huggingface.co) · 切换

下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。

权重 koichame_krea2_v1_bf16.safetensors 224 MB 下载
权重 koichame_krea2_v1_fp32.safetensors 448 MB 下载

说明

Koichame Style — Dreamy Pastel Anime

作者:Mike_832

Koichame Style — Dreamy Pastel Anime A style LoRA based on illustrations by koichame. Think coloured linework, soft pastels and flowing hair and fabric. I like it on character illustrations with a strong shape behind the face, such as roses against a dark background or a full moon over water.

  • Krea 2 — @koichame
  • Anima — @koichame

About this version — Krea 2 Trigger @koichame. including the period. Describe the scene in plain sentences. Prompt template @koichame. <subject and appearance>. <action and setting>. <light and mood>. LoRA strength 1.0 Recommended settings

  • Krea 2 Turbo FP8 · 832 × 1248 · Euler / simple · 8 steps · CFG 1
  • The ComfyUI workflow uses ConditioningZeroOut for the negative conditioning.
  • For a larger image, resize to 1216 × 1824 with Lanczos and run another 8 steps at CFG 1. Use the same prompt, seed and LoRA strength. Try denoise 0.15 or 0.25.

Check the eyes after hires. Both settings can redraw the eyelashes and irises. I kept the first pass on the cover and several close-ups. Tips

  • For a pair, give each character their own hair, eye and clothing details.
  • The LoRA can apply the style even without the trigger at strength 1.0. Some character traits can carry across, such as pointed ears on Fern, and everyday scenes can occasionally come out photographic.

Epoch 12 of 12 at 1,560 steps. Training and files Trained on Krea 2 Raw with Musubi Tuner using 65 illustrations with long and short natural-language captions. Rank and alpha are both 32. Choose the original FP32 weights or BF16 at about half the file size. ────────────────────────────────────────────────────── About this version — Anima Base model: Anima (anima-base-v1.0) Trigger: @koichame (at the very start of the prompt, followed by a comma, then regular Danbooru tags). It is the artist's own Danbooru tag; base Anima knows it a little, the LoRA makes the style far softer and calmer than the tag alone. No style tags needed. LoRA strength: 1.0 Hires fix is optional here and needs a low denoise (see below). Recommended settings

  • Sampler: DPM++ 2M SDE, scheduler simple, 25 steps, CFG 5
  • Size: about 1 megapixel in any aspect ratio: 832×1216, 1024×1024, 1216×832, 1344×768 all work.
  • Hires fix (optional): lanczos 1.5×, 20 steps, denoise 0.20, with the LoRA still loaded at 1.0.

   At 0.20 the hires pass adds a little depth to hair, lace and jewellery. Higher denoise (0.35) adds sparkles and hard lines and redraws the eyes, and even at 0.20 it can change eye colour or the shape of closed eyes, and redraws small faces in full-body shots. Check the eyes: if the hires pass spoils them, keep the first pass, it is clean on its own. Negative prompt: worst quality, low quality, score_1, score_2, score_3, jpeg artifacts, logo, watermark, signature, twitter username, bad anatomy, bad hands, missing finger, extra digits, fewer digits, nsfw, nude, explicit Tips

  • Describe the scene with flowers, water, sky or a moon. That is where the style shines: bouquets, roses, petals, lily pads, night sky, clouds.
  • Keep "signature" in the negative. The artist signs most works and Anima associates the tag with a small handwritten signature; with the negative it is rare, but a faint one can still show up.

Known issues

  • Very light, airy pieces come out a little darker and greyer than the artist's own: the LoRA keeps the soft colours but less of the glowing white.
  • Hires fix can damage eyes even at low denoise (see above).
  • Small text and numbers (Miku's "01" tattoo, signs) can be smeared.
  • Trained on Anima Base v1.0 only. Behaviour on Anima Turbo and third-party merges is unknown.
  • Showcase images: prompts made from the tags of the artist's own works, original aspect ratios at about 1 MP; 23 of 33 with hires 1.5× at denoise 0.20, 10 kept as the first pass where hires spoiled the eyes. Full prompts are in the image metadata.

Training 65 images out of 75: exact duplicates and smaller copies of the same artwork removed, three tiny eye close-ups (under 700 px) removed; signed images kept, with the signature tagged so it can be pushed into the negative; portrait, square and landscape mixed Short tag captions: trigger + rating + people count + character names + content tags from WD EVA02 Large v3 (names spot-checked by hand); nothing about style, palette, lighting or medium, no quality tags Anima LoRA, rank 64 / alpha 64, incl. AdaLN modulation; DiT trained, text encoder and LLM adapter frozen · AdamW8bit, LR 1e-4, constant with 5% warmup · batch 1 · 1024 res, buckets 512–1536 · sigmoid timesteps, scale 1.3 · caption dropout 5% · 12 epochs Release file: epoch 8 of 12 (520 steps). Epoch 10 fell into a template (flat grey ground, tiny figures, objects dropped from the prompt); epoch 12 drew everything brighter than the artist's own works Picked from all epochs 2–12 on 6 scenes, then an overfitting check on 6 prompts far from the dataset × 3 seeds, then a comparison with 12 of the artist's own works in their original aspect ratios

示例图