Anima29B P1 + Turbo 1.1 Merged Model (Unofficial Merge)

Unofficial community derivative: a full UNET produced by block-merging Anima 2.9B Preview 1 with Anima Turbo 1.1. NOT an official release by Anima / CircleStone Labs.

Use directly: load it as the diffusion model — no LoRA or patch nodes needed.

• File: anima29b-p1-turbo1.1.safetensors

• Size: ~5.4 GiB — full model (merged UNET, 40 blocks, BF16)

Goal: turbo speed (8 steps) with Anima 2.9B Preview 1-tier quality, character/artist knowledge and NSFW performance.

How it was merged

Shared blocks 0.8 × Turbo 1.1 + 0.2 × 2.9B P1; the 12 extension blocks unique to 2.9B P1 are kept as-is. The 0.8 strength was settled by full-scene same-seed comparisons: SFW and 100%-crop details on par across ratios, 0.8 slightly better in NSFW anatomy; no spontaneous mosaic / black bars at any ratio.

Workflow

UNETLoader (this model) + CLIPLoader (qwen_3_06b_base.safetensors) + VAELoader (qwen_image_vae.safetensors) → two CLIPTextEncode (positive/negative) → KSampler → VAEDecode → SaveImage. No LoRA needed.

Recommended parameters

• steps 8 (8-12)

• sampler euler + sgm_uniform

• cfg: 1.0 without negative; 1.2 with negative (range 1.0-1.6 — negatives take effect as soon as cfg > 1)

• resolution: 1024×1024 baseline; 1264×848 landscape / 848×1264 portrait tested fine (~14-17 s/image)

Prompting (quality tags work)

Same as the turbo family — masterpiece, best quality etc. work:

masterpiece, best quality, <safe/sensitive/nsfw/explicit>, <count>, <character (series)>, @<artist>, <appearance tags>, <natural language sentences>

• Character name right after the series tag; avoid underscores (hatsune miku, not hatsune_miku)

• @artist strongly steers the style — recommended

• Write 1-2 full natural-language sentences; for multiple characters describe each one's appearance next to them

• Standing text-type negative text, letters, sign, logo, watermark, signature, username, artist name + cfg 1.2 recommended: noticeably reduces corner fake signatures / watermarks / logos (useless against text that belongs to the scene itself); for scenes without an artist tag, an empty negative + cfg 1.0 is fine

• Not good at rendering readable text; for NSFW just write the rating and content tags directly

Sample images

All samples (s1/s3/s4/s5) are straight outputs of this model at 8 steps, cfg 1.2 with the text-type negative above. Full ComfyUI workflow metadata is embedded in the PNGs — drop into ComfyUI to reproduce.

Related

• Sister project (LoRA version): Anima29B Turbo v2.9 LoRA (Unofficial Remap) (anima29b-turbo-v2.9-remap) — same speed on the original 2.9B base without swapping the checkpoint

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

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

License

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

Anima29B P1 + Turbo 1.1 合并模型（非官方 Unofficial Merge）

非官方（unofficial）社区衍生模型：由 Anima 2.9B Preview 1 与 Anima Turbo 1.1 经块级合并（block-merge）得到的完整 UNET，非 Anima / CircleStone Labs 官方发布。

直接用：作为 diffusion model 加载即可，无需任何 LoRA 或补丁。

• 文件：anima29b-p1-turbo1.1.safetensors

• 大小：约 5.4 GiB —— 完整模型（合并版 UNET，40 块 BF16）

目标：turbo 速度（8 步）+ Anima 2.9B Preview 1 档画质、角色/画师知识与 NSFW 表现。

合并方式

共享块 0.8 × Turbo 1.1 + 0.2 × 2.9B P1，2.9B 独有的 12 个扩展块原值保留。0.8 为全场景横评定案：SFW 各配比同档、100% 放大细节同档，NSFW 解剖细节 0.8 略优；任何配比均无自发马赛克/黑条。

工作流

UNETLoader（本模型）+ CLIPLoader（qwen_3_06b_base.safetensors）+ VAELoader（qwen_image_vae.safetensors）→ 双 CLIPTextEncode（正/负）→ KSampler → VAEDecode → SaveImage。无需 LoRA。

推荐参数

• 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，cfg>1 负面即生效）

• 分辨率：1024×1024 基准，横 1264×848 / 竖 848×1264 实测良好（约 14-17 秒/张）

提示词（质量词有效）

与 turbo 系一致，masterpiece, best quality 等质量词有效：

masterpiece, best quality, <safe/sensitive/nsfw/explicit>, <人数>, <角色 (作品)>, @<画师>, <外貌标签>, <自然语言句子>

• 角色名紧跟作品标签，标签避免下划线（写 hatsune miku 而非 hatsune_miku）

• @画师 显著影响画风，建议显式给出

• 自然语言写足 1-2 句；多角色逐角色就近描述外貌

• 建议常驻文字类负面 text, letters, sign, logo, watermark, signature, username, artist name + cfg 1.2，可显著减少角落假签名/水印/logo（对画面内容本身的文字无效）；不带画师场景可留空负面 + cfg 1.0

• 模型不擅长生成可读文字；NSFW 直接写分级与内容标签即可

示例图

4 张示例（s1/s3/s4/s5）均为本模型 8 步、cfg 1.2 + 上述文字类负面直出。PNG 内含完整 ComfyUI 工作流，拖入即可复现。

关联推荐

• 姊妹项目（LoRA 版）：Anima29B Turbo v2.9 LoRA（非官方块重映射版）（anima29b-turbo-v2.9-remap）——不想换底模时用，配合原版 Anima-2.9B Preview 1 生效

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

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

License

上游 Anima / Anima-2.9B / Turbo-ANIMA 均采用 CircleStone Labs Non-Commercial License，本合并模型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。