EVERY SINGLE GENERATION PUTS ON WEDDING RING!!! MAKE IT STOP!!!
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 25/100
- Issue 类型
- 缺陷
- 描述清晰度
- 需要澄清
- 活跃度
- 活跃
- 技术栈
- python
- 领域
- ai, machine-learning
调研方向
The issue names no files, tests, or code entry points. Start by reviewing the available prompts, character and hand references, failed outputs, API task IDs, and local configuration the reporter can provide; compare API and local Ref2VA results. Done means determining whether the behavior can be reproduced and whether this repository has an actionable fix.
由索引模型根据 Issue 内容生成。
描述
Repeated hallucination of wedding rings on characters with bare-hand references
Summary
MiniMax H3 repeatedly generates wedding rings or ring-like jewelry on characters whose reference images clearly show bare hands.
This happens both through the official MiniMax API and through local/open-weight H3 Ref2VA workflows.
Expected behavior
If the supplied character references and dedicated hand references show bare fingers, H3 should preserve that detail and not invent rings or other finger jewelry.
Actual behavior
H3 frequently adds a ring, usually resembling a wedding band, even when:
- The character reference shows bare hands.
- Dedicated hand-reference images also show bare fingers.
- The prompt contains no request for jewelry.
- The intended character design has no rings.
The issue becomes especially noticeable when the hands are visible or interacting with objects.
Reproducibility
have reproduced this behavior with:
- The official MiniMax H3 API.
- Local H3 Ref2VA generation.
Because it occurs in both environments, this does not appear to be specific to Wan2GP, ComfyUI, or another frontend.
Why this matters
This is a character-continuity problem.
A wedding ring is not a minor texture variation. It changes a visible character attribute and can make an otherwise usable shot unusable in a sequence.
In my case, shots have needed to be regenerated solely because H3 invented rings that do not exist in the character references.
Possible cause
This appears to be a reference-fidelity or model-prior issue where learned visual priors are overriding details established by the references.
It may be worth investigating whether hand details and the absence of accessories are being underweighted during Ref2VA conditioning.
Additional material available
I can provide:
- Exact prompts
- Character reference sheets
- Dedicated hand references
- Failed outputs
- API task IDs
- Local configuration details
- Multiple examples showing the same issue across different shots
- 主要语言
- Python
- 星标
- 9.8k
- 派生
- 738
- PR 合并指标
- 30 天内没有已合并 PR
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
MiniMax-AI/MiniMax-H3 的其他 Issue
-
loss of init.py未关闭
难度 3/5 半天 新手友好度 15/100
MiniMax-AI/MiniMax-H3#92 · 1 条评论 ·
-
难度 5/5 一周以上 新手友好度 5/100
MiniMax-AI/MiniMax-H3#89 ·
-
难度 5/5 一周以上 新手友好度 5/100
MiniMax-AI/MiniMax-H3#88 ·
-
،her未关闭
难度 5/5 一周以上 新手友好度 10/100
MiniMax-AI/MiniMax-H3#87 ·
-
难度 3/5 1-2 天 新手友好度 74/100
MiniMax-AI/MiniMax-H3#85 ·
查看 MiniMax-AI/MiniMax-H3 的全部 Issue
相似的 Issue
-
good first issue
难度 2/5 1-3 小时 新手友好度 85/100
btclib-org/btclib-node#1955 ·
维护者通常 1 天内回复
-
documentation v0.9.2
难度 2/5 半天 新手友好度 72/100
khuisman/mcp-gee-sweet#1033 ·
维护者通常 1 天内回复
-
bug
难度 2/5 1-3 小时 新手友好度 76/100
ossf/cve-bin-tool#5966 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 62/100
维护者通常 2 天内回复
-
write_iceberg ignores the table's write.parquet.compression-codec property可能已有人在做 @YuangGao 今天认领。 未关闭
难度 2/5 1-3 小时 新手友好度 82/100
Eventual-Inc/Daft#7606 ·
维护者通常 1 天内回复