featuring:A universal FL2VA / REF2VA hybrid base modelA Low-Sigma second-pass workflowMain goals: reduce noise, recover detail, and improve overall visual qualityThe second-pass workflow is based on a segmented denoising approach using Sigma Extension + SplitSigmasDenoise.By splitting the H3 denoising process with Split Sigma, the workflow separates it into two stages:Stage 1 — Motion GenerationUses relatively higher Turbo / motion LoRA weights to handle movement, temporal structure, and action generation.Stage 2 — Low-Sigma Detail RefinementNo additional noise is introduced, while LoRA strength is reduced. Only the remaining low-Sigma section is used to continue refining the existing latent.The goal is to reduce noise and recover finer details while preserving the original motion and temporal logic as much as possible.About the Base ModelThe workflow uses the recommended hybrid model:<a target="_blank" rel="ugc" href="https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/tree/main">minimax_h3_hybrid_fl2va_ref2va_b25-49</a>Its basic principle is to use FL2VA as the main model, while replacing only the adaln_proj conditioning modulation layers of selected Transformer Blocks with weights from Ref2VA.This means there is no need for a dual-model second pass. A single hybrid model can attempt to combine:FL2VA image and audio qualityRef2VA reference consistencyThe name b25-49 means that Transformer Blocks 25–49 use the Ref2VA modulation-layer weights.Differences Between the Hybrid VariantsThe main difference is how many later Transformer Blocks use Ref2VA adaln_proj weights.Generally, using more Ref2VA blocks provides stronger reference consistency, but image and audio characteristics gradually move closer to Ref2VA.b30-49More FL2VA-oriented. Best image/audio quality, but relatively weaker reference consistency.b25-49Balanced version. Good compromise between image quality and reference consistency, recommended as the default.b20-49More Ref2VA-oriented. Stronger reference consistency, with a slight reduction in image/audio quality.b15-49The most Ref2VA-oriented version. Strongest reference consistency, but also the largest quality trade-off compared with the original FL2VA model.Overall, this setup supports both FL2VA and REF2VA workflows without maintaining two separate base models, and can be used on 12GB GPUs.<hr />本期分享一套 MiniMax H3 12GB 可用方案：FL2VA / REF2VA 通用底模低 Sigma 二采工作流核心作用：去噪点 + 补细节 + 提升观感这套二采基于Sigma 扩展 + SplitSigmasDenoise 的分段思路：通过 Split Sigma 将 H3 去噪过程拆成动作生成阶段与低 Sigma 细节收敛阶段。第一阶段保持较高 Turbo / 动作 LoRA 权重负责运动和时序，第二阶段关闭新增噪声并降低 LoRA 权重，只利用末段低 Sigma 对已有 latent 继续收敛，从而在尽量不改变原有动作逻辑的情况下减少噪点并补充细节。关于底模：使用的别人推荐的minimax_h3_hybrid_fl2va_ref2va_b25-49原理：以 FL2VA 为主体，仅将部分 Transformer Block 的 adaln_proj 条件调制层替换为 Ref2VA 权重，不需要双模型二采，单模型即可尝试兼顾 FL2VA 的画质与 Ref2VA 的参考能力。b25-49 即第 25–49 Block 使用 Ref2VA 的调制层权重。b30-49：更偏 FL2VA，画质/音频质量最好，参考能力相对弱一些。b25-49：偏均衡，画质和参考能力兼顾，适合作为默认版本。b20-49：更偏 Ref2VA，参考一致性更强，但画质/音频略有下降。b15-49：最偏 Ref2VA，参考能力最强，同时也是四个版本里对原始 FL2VA 质量牺牲最大的。整体方案支持 FL2VA / REF2VA 通用，不需要分别维护两套底模，12GB 显卡可用。

Files with SHA256: 119a1a972714c2d3fffd6c957492e1080d80b01821fd8a45050279a3263a7554