Anima29B Turbo v2.9 LoRA (Unofficial Block Remap)

Unofficial community derivative LoRA: a block-index remapped version of Turbo-ANIMA-v2.9 for Anima-2.9B. NOT an official release by Anima / CircleStone Labs.

Use with the original Anima-2.9B Preview 1: gives the 2.9B base turbo speed (8 steps) while keeping its quality, character/artist knowledge and NSFW performance.

• File: anima29b-turbo-v2.9-remap.safetensors

• Size: ~227 MiB — LoRA (written directly in the 40-block layout)

Why the remap

Anima 2.9B is a 28→40 block depth extension of Anima base (new blocks inserted at indices 2,5,8,11,14,17,21,24,27,30,33,36). The original Turbo-ANIMA LoRA is in the 28-block layout and silently misaligns when attached to 2.9B (a runtime patch is otherwise required). This remapped version has its keys rewritten offline to the 40-block layout — it loads correctly without any runtime patch. Verified pixel-identical against "original LoRA + runtime patch" under the same seeds (6 scenes incl. NSFW).

Workflow

UNETLoader (anima29B_v10.safetensors, the original Anima-2.9B Preview 1) → LoraLoaderModelOnly (this LoRA, strength 0.8) + CLIPLoader (qwen_3_06b_base.safetensors) + VAELoader (qwen_image_vae.safetensors) → two CLIPTextEncode → KSampler → VAEDecode → SaveImage.

Recommended parameters

• LoRA strength 0.8 (1.0 usable; 0.5 under-converged, unusable)

• steps 8 (8-12)

• sampler euler + sgm_uniform

• cfg: 1.0 without negative; 1.2 with negative (range 1.0-1.6)

• resolution: official 812×1216 / 1152×1536 / 1536×1536; 16-multiples like 848×1264 tested fine (~11-19 s/image)

Prompting (NO quality tags)

The 2.9B base has no quality/score tags in its training data — do not use masterpiece / best quality / score_N:

year 2026, <safe/sensitive/nsfw/explicit>, <count>, <character (series)>, @<artist>, <appearance tags>, <natural language sentences (2-3+)>

• Keep the year 2026 tag; character right after series; avoid underscores

• An explicit @artist tag is recommended to anchor the style; write 2-3+ full natural-language sentences

• Negative generally stays empty (cfg 1.0); against fake signatures/watermarks write text, letters, sign, logo, watermark, signature, username, artist name + cfg 1.2

• For NSFW just write the rating and content tags; no "uncensored"-style positive tags needed

Sample images

Samples are straight outputs of this LoRA at strength 0.8, 8 steps, cfg 1.0. Full ComfyUI workflow metadata is embedded in the PNGs — drop into ComfyUI to reproduce.

Related

• Sister project (full model): Anima29B P1 + Turbo 1.1 Merged Model (anima29b-p1-turbo1.1) — no LoRA needed, quality tags work, turbo prompt style

• ComfyUI-AnimaFlow: AI tool-use API (MCP / HTTP) for Anima-family models, ships a ready-made workflow (anima29_turbo) and prompt knowledge for this LoRA

• Upstreams: Turbo-ANIMA, Gazingstars123/Anima-2.9B, circlestone-labs/Anima

License

All upstreams (Turbo-ANIMA / Anima-2.9B / Anima) use the CircleStone Labs Non-Commercial License; as a derivative work this remapped LoRA follows the same license: non-commercial use only, and the license notice must be preserved on redistribution.

Anima 2.9B Turbo v2.9 LoRA（非官方块重映射版 Unofficial Remap）

非官方（unofficial）社区衍生 LoRA：Turbo-ANIMA-v2.9 的 Anima-2.9B 块索引重映射（remap）版本，非 Anima / CircleStone Labs 官方发布。

配合原版 Anima-2.9B Preview 1 使用：给 2.9B 底座加上 turbo 速度（8 步），底座模型的画质、角色/画师知识与 NSFW 表现不变。

• 文件：anima29b-turbo-v2.9-remap.safetensors

• 大小：约 227 MiB —— LoRA（40 块布局直写）

为什么需要重映射版

Anima 2.9B 是 base 的 28→40 块深度扩展（新块插入在索引 2,5,8,11,14,17,21,24,27,30,33,36）。原版 Turbo-ANIMA LoRA 是 28 块布局，直接挂到 2.9B 上键索引会静默错位（需要运行时补丁动态修正）。本重映射版已离线把键索引改写为 40 块布局，无需任何运行时补丁即可正确加载生效——已与"原版 LoRA + 运行时补丁"做同 seed 逐像素对照（6 场景含 NSFW），输出完全一致。

工作流

UNETLoader（anima29B_v10.safetensors，即原版 Anima-2.9B Preview 1）→ LoraLoaderModelOnly（本 LoRA，strength 0.8）+ CLIPLoader（qwen_3_06b_base.safetensors）+ VAELoader（qwen_image_vae.safetensors）→ 双 CLIPTextEncode → KSampler → VAEDecode → SaveImage。

推荐参数

• LoRA 强度 0.8（1.0 可用；0.5 欠收敛不可用）

• steps 8（可用 8-12）

• 采样器 euler + sgm_uniform

• cfg：不写负面 1.0；写负面 1.2（范围 1.0-1.6）

• 分辨率：官方推荐 812×1216 / 1152×1536 / 1536×1536，848×1264 等 16 倍数实测良好（约 11-19 秒/张）

提示词（不要质量词）

2.9B 底座训练数据无质量/分数标签，禁用 masterpiece / best quality / score_N 类标签：

year 2026, <safe/sensitive/nsfw/explicit>, <人数>, <角色 (作品)>, @<画师>, <外貌标签>, <自然语言句子 2-3 句>

• 保留 year 2026 年份标签；角色名紧跟作品标签，避免下划线

• 建议显式给 @画师 锚定画风；自然语言写足 2-3 句

• 负面一般留空（cfg 1.0）；排除假签名/水印时写文字类负面 text, letters, sign, logo, watermark, signature, username, artist name 并配 cfg 1.2

• NSFW 直接写分级与内容标签，无需额外正面"无码"类提示词

示例图

示例为 LoRA 版（强度 0.8、8 步、cfg 1.0）直出。PNG 内含完整 ComfyUI 工作流，拖入即可复现。

关联推荐

• 姊妹项目（完整模型版）：Anima29B P1 + Turbo 1.1 合并模型（anima29b-p1-turbo1.1）——不想挂 LoRA 时直接用，质量词有效的 turbo 提示词体系

• ComfyUI-AnimaFlow：Anima 系模型的 AI 工具调用 API（MCP / HTTP），内含本 LoRA 对应的开箱工作流（anima29_turbo）与提示词知识

• 上游：Turbo-ANIMA、Gazingstars123/Anima-2.9B、circlestone-labs/Anima

License

上游 Turbo-ANIMA / Anima-2.9B / Anima 均采用 CircleStone Labs Non-Commercial License，本重映射 LoRA 作为衍生作品遵循同一许可：仅限非商业用途，再分发需保留许可声明。