Minimax-H3 Subtitle-Control LoRA (Experimental)

This repository contains an experimental LoRA checkpoint specifically customized and fine-tuned for the Minimax-H3 Hybrid model (combining FL2VA front 25 layers + REF back 25 layers).

It is designed to solve subtitle leakage issues and restore strong prompt controllability over burned-in text, while simultaneously improving instruction adherence and multi-parameter generation stability.

🎬 Visual Showcase (Side-by-Side Comparison)

Here is a 12-second generation result demonstrating multi-turn rapid dialogue with zero subtitle leak, high lip-sync fidelity, and strong identity retention from the reference image:

📸 Reference Image (`<Picture 1>`)	🎥 Generated Video (LoRA Scale 1.0, Subtitle-Free)
Character Identity & Attire Reference Refined features, beauty mark & period costume	12.0s · 768×1344 (9:16) · 8 Steps · Singularity + vdn-dmd Fused Inference (last 10 layers weight: 0.7) 8-turn rapid dialogue with zero subtitle leak & clean lip-sync

📜 Click to view complete 12s generation prompt

subject_definitions:
<Subject 1> is the young woman in <Picture 1>, with refined features, soft dark hair styling, and a distinct beauty mark near her left eye, wearing realistic ancient Chinese silk robes.
<Subject 2> is the young man, dressed in realistic ancient Chinese robes, with a sincere and earnest expression.
<Audio 1> is the ambient evening soundscape, including subtle wind and distant crickets.

summary:
[reference generation] The target video is a 12-second realistic ancient Chinese period drama scene featuring a rapid-fire dialogue exchange between <Subject 1> and <Subject 2> in a lantern-lit courtyard corridor at dusk, concluding with a comedic reaction shot.

retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - the appearance, hair, and beauty mark from <Picture 1> are retained.
<Subject 2> (appears in [Shot 1], [Shot 2]): fully_preserved - the sincere expression and period-accurate attire are retained.
<Audio 1>: fully_preserved - the ambient evening soundscape is used throughout the scene.

detailed_description:
The target video is in a realistic, cinematic ancient Chinese period drama style, with natural dusk lighting and warm, soft lantern light highlighting the intricate textures of silk robes and skin.
[Shot 1] The scene opens in an ancient courtyard corridor at dusk. <Subject 1> (S1) and <Subject 2> (S2) stand facing each other at a natural conversational distance. <Subject 2> (S2) asks tentatively, <d>[Chinese] 晚上出去喝点吗？</d> <Subject 1> (S1) raises an eyebrow and asks, <d>[Chinese] 我喝醉了你会不会动手动脚？</d> <Subject 2> (S2) replies earnestly, <d>[Chinese] 不会。</d> <Subject 1> (S1) continues, <d>[Chinese] 会不会抱我？</d> <Subject 2> (S2) answers, <d>[Chinese] 不会。</d> <Subject 1> (S1) asks, <d>[Chinese] 会不会欺负我？</d> <Subject 2> (S2) says firmly, <d>[Chinese] 绝对不会！</d> <Subject 1> (S1) stares at him for a half-second, then says with a mix of disdain and amusement, <d>[Chinese] 这也不会那也不会，那还喝个锤子啊？</d>
[Shot 2] At 10.000, the shot cuts to a tight close-up of <Subject 2> (S2). He stands frozen with a bewildered, stunned expression, his eyes wide as he processes the punchline.

overall_soundscape:
<Audio 1> provides a consistent ambient evening soundscape of subtle wind and distant crickets throughout the video.

non_diegetic_music:
None.

📥 Quick Download & Placement

For users running custom pipelines or ComfyUI:

Weights File: contextir-hybird-v0.1.safetensors
Recommended Placement: ComfyUI/models/loras/ or your framework's LoRA directory

Direct Download Command:

wget -O contextir-hybird-v0.1.safetensors https://huggingface.co/Justlora/MinimaxH3-Subtitle-Control-LoRA/resolve/main/contextir-hybird-v0.1.safetensors

💡 Background & Motivation

In hybrid inference configurations of Minimax-H3 — specifically the Hybrid FL2VA (first 25 layers) + REF (last 25 layers) fusion mode commonly used in high-performance serving environments (such as WarmBloodAban/Minimax-h3_Singularity) — community users frequently encounter a notable drawback:

Degraded Subtitle Controllability: Under certain scenarios and complex multi-parameter settings, standard negative prompts (e.g., subtitles, text, watermark) lose their effectiveness, causing unwanted subtitles and burned-in text to persist in the generated video frames (as documented and discussed in Singularity Discussion #24: Subtitle problem).

This LoRA is tailored to address this root issue by re-aligning attention weights across both the early representation (FL) and refinement (REF) stages of the DiT backbone.

🌟 Key Mechanism: Quoteless `<d>` Dialogue Formatting

With this LoRA applied, the model supports clean dialogue control using the <d>[Language] content </d> syntax:

Crucial Rule: Ensure dialogue inside <d>...</d> tags contains NO quotation marks (neither ASCII " nor Chinese full-width “”).
When quotes are omitted, the model reliably treats the text solely as audio/dialogue guidance without visualizing or burning text onto video frames.
Drastically Reduced Reroll Rate: Eliminates repetitive failed generations ("gacha rate"), significantly boosting generation consistency and success rate in production workflows.

📝 Prompt Example Comparison

✅ Recommended (Guaranteed Subtitle-Free)

[Shot 1] The young man looks over and asks tentatively<d>[Chinese] 晚上出去喝点吗？</d>

❌ Avoid (Prone to Burned-in Subtitles)

[Shot 1] The young man looks over and asks tentatively: "晚上出去喝点吗？" (Quotation marks easily trigger on-screen subtitle rendering)

⚙️ Model Specifications

Attribute	Specification
Base Model	Minimax-H3 Hybrid Model (FL 1–25 + REF 26–50 Fusion)
Upstream Architecture	MiniMax-H3 (`MiniMaxAI/MiniMax-H3`)
Weight Format	SafeTensors (`contextir-hybird-v0.1.safetensors`)
Precision / Dtype	BF16 (`bfloat16`)
LoRA Rank	16
File Size	~65.6 MB
Target Attention Layers	`blocks.0` ~ `blocks.51` (`attn.qkv_proj`, `attn.out_proj`)
Checkpoint Status	Experimental

🚀 Recommended Inference Configuration

Model & Pipeline Compatibility:
- This LoRA is trained natively on the Minimax-H3 Hybrid model (FL 1–25 + REF 26–50 fusion).
- It can be directly mounted and used on Singularity pipelines as well as Hybrid Minimax-H3 serving frameworks.
Recommended Distillation / Schedulers:
- Singularity + vdn-dmd fused inference (last 10 layers weight: 0.7) / lavarryv4turbo
Inference Steps:
- 8 steps
Recommended LoRA Scale:
- 1.0

📄 License & Compliance

LoRA Adapter Weights: Released under the permissive Apache 2.0 License.
Base Model Compliance: The underlying Minimax-H3 base model is subject to the MiniMax H3 Community License Agreement.

⚠️ Disclaimer

This checkpoint is an experimental exploration release. It significantly mitigates subtitle leakage and improves instruction adherence across most scenarios. If you encounter any bugs, community feedback and discussions are warmly welcomed! We will actively follow up and address them.

Downloads last month: -

Model tree for Justlora/MinimaxH3-Subtitle-Control-LoRA

Base model

MiniMaxAI/MiniMax-H3

Adapter

(158)

this model