触发词
文件清单 线路:镜像(hf-mirror.com) · 切换
下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。
说明
Vesperveil — Painterly Oil & Gouache
作者:Mike_832
Vesperveil — Painterly Oil & Gouache Original painterly style LoRA for CircleStone Anima (anima-base-v1.0). Rich oil and gouache brushwork with matte impasto texture, expressive lighting, and atmospheric depth — turning anime characters, fantasy scenes, and full environments into cohesive hand-painted illustrations. Works equally well on portraits and wide scenic compositions: moonlit ruins, neon rain, sunset skies, lantern-lit streets, medieval castles, and quiet interiors. Trained on a color-balanced dataset, so it respects both cool nocturnal palettes and warm golden light without forcing a single tint. ────────────────────────────────────────────────────────── About this version — Anima Base model: Anima (anima-base-v1.0) Trigger: @vesperveil (invented trigger; include it near the start of the prompt) No style tags needed. You do not need to prompt "oil painting" or "gouache". LoRA strength: 1.0 (or 0.8 for slightly softer faces; below 0.75 the brushwork fades) Hires fix is part of the recipe, not an option (see recommended settings below). Recommended settings
- Sampler: DPM++ 2M SDE (dpmpp_2m_sde_gpu), scheduler simple, 20–25 steps, CFG 5
- Size: ~1 megapixel in portrait or landscape — 832×1248 (2:3), 1184×896 (4:3), 1248×832 (3:2), or 1344×768 (16:9).
- Hires fix: Lanczos 1.5× (e.g. 1248×1872 or 1776×1344), 20 steps, denoise 0.35 (0.30–0.35 works well), with the LoRA loaded at 1.0 on BOTH passes.
The first pass builds composition and silhouettes; the hires pass at denoise 0.35 is where the thick gouache brushwork, canvas depth, and painterly edges really come alive. Keep the LoRA active during hires (otherwise base Anima smooths the texture back into flat digital anime), and avoid upscaling past 1.5× on Anima. Negative prompt: worst quality, low quality, lowres, artist name, signature, text, (bad anatomy:1.2), (bad hands:1.1), missing finger, (extra digits:1.4), fewer digits, disfigured, mutation, 4 fingers, (6 fingers:1.4) Tips & Face Style Control
- Want softer anime faces instead of semi-realistic proportions? Because the training set is semi-realistic painting, adult characters (such as Makima or Tifa) can occasionally get sharper jawlines or heavier facial structure. To steer faces back toward classic anime proportions while keeping full oil/gouache brushwork at LoRA 1.0:
– Add to positive: anime style, anime face – Add to negative: realistic, photorealistic, semi-realistic, sharp jawline, masculine Combining both gives the cleanest result (the negative alone is subtle). You can also lower LoRA strength to 0.8 for a softer look, but dropping to 0.7 noticeably flattens the brushwork.
- Prompt lighting, framing, and environment explicitly. Tags like dramatic backlighting, warm pendant lamps, moonlight, low angle, wide shot, detailed background unlock the full painterly depth of the model.
────────────────────────────────────────────────────────── Known issues
- Adult characters can occasionally develop sharp jawlines or overly sculpted faces on default prompts — use the anime-face tip above whenever you want softer features.
- On complex props (weapons, vehicles, staves, railings), check the first pass for physical continuity before running hires.
- Trained on Anima Base v1.0 only. Behavior on Anima Turbo or third-party merges is untested.
- Showcase images use 1.5× Lanczos hires at denoise 0.35 with the LoRA at 1.0 on both passes. Full prompts and ComfyUI graphs are embedded in the PNG metadata.
────────────────────────────────────────────────────────── Training 155 curated painterly oil & gouache images (balanced across cool, neutral, and warm color temperatures; anime-proportioned faces excluded from the training set so brushwork stays consistent; portrait, square, and landscape mixed) Short minimal captions (median 8 tags, range 6–12): @vesperveil + rating + people count + core subject and setting tags (manually reviewed); no quality tags, no style, medium, palette, or lighting 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 (1860 steps) Release file: epoch 8 of 12 (1240 steps). Later epochs (11–12) pushed facial structure too hard; epoch 8 balances rich painterly texture with clean likeness and prompt adherence Evaluated across epochs 2–12 on 16+ scenes, followed by a 24-scene stress test, hires denoise comparisons (0.25 / 0.30 / 0.35) paired against base Anima, and a controlled 6-variant anime-face steering test
