MiniMax H3 Evening Report 2026-11-04 | First Recommendation: MiniMax H3_SparseRef15_Hybrid

Zhenzhen's AI Radar · MiniMax H3 · 2026-11-04 Evening Report

This issue features 7 new versions: Strongly Recommended 0 · Recommended 1 · Worth Watching 4 · Listed Only 2

Editor's Note This issue reviews 7 new versions of H3. The top recommendation is the Turbo INT8 version of SparseRef15 Hybrid: it sparsely injects Ref2VA reference capabilities into 15 blocks of the FL2VA base model. The author claims that character consistency remains high during multi-segment long video generation. This version produces output in 4 steps, but only 1 visible examples are provided and no sampler is specified, earning it an A rating. The B tier includes a video replacement workflow that swaps actors using a single photo, a workflow v0.2 that uses H3 as a single-frame image editor, a paid snow slider with positive and negative weights, and a tail concept LoRA. The fictional character with all examples hidden and the The Eminence in Shadow style LoRA, which the author is dissatisfied with, are rated C.

Recommended (1)

[A] MiniMax H3_SparseRef15_Hybrid · Turbo_PP_INT8 V1.0

Base Model / Checkpoint · Version Update · Author Aki7777777 · Published 11-04 18:20 · Downloads 133

The INT8 version of the 4-step Turbo part of SparseRef15 Hybrid: Based on FL2VA, it sparsely adds Ref2VA reference capabilities to 15 Transformer blocks. The goal is to prevent character drift and maintain free movement during multi-segment long video generation.

Prioritize this when the same character needs to appear in many consecutive segments; comparing it with the original FL2VA using the same reference image for multi-segment generation best highlights the differences.

Reasons for Recommendation

• Version Update: This is the 4-step version. The author notes that videos with large motion amplitudes are recommended to use 6 steps or more; the series previously had multiple pruned and quantized precision versions.
• Approach: Instead of merging LoRAs, it only uses AdaLN to inject Ref2VA into 15 blocks within the 50 block backbone. The author claims this preserves FL2VA's motion freedom and scene representation while enhancing character identity consistency.
• Serves long videos requiring continuous multi-segment generation of the same character; the author claims drift is minimal during character turns, temporary absence from the front view, and consecutive multi-segment scenarios.
• Partial INT8: Large matrices are converted to INT8 ConvRot, while AdaLN and time conditions retain FP32. The author claims this version is faster and saves more VRAM than the non-INT8 version; this version has 133 downloads, and the model has accumulated 9696 downloads (indicating popularity only).

How to Use

• Download minimaxH3Sparseref15_turboPPINT8V10.safetensors (INT8) to replace the H3 base model. Use 4 steps; use 6 steps or more for large motion shots.
• It is not a native Ref2V model. The author requires LoRAs compatible with FL2VA; do not use Ref2V LoRAs.
• The sampler, CFG, reference image connection method, and workflow are not provided by the author.

Risks and Limitations

• Only 1 video segments are visible as examples (another 2 sensitive-level examples are hidden); "almost no drift in long videos" is the author's claim, with no comparative examples.
• The author claims that branches with stronger reference capabilities sometimes compress scene freedom, and scenes where characters need to be off-screen for a long time may be "pulled back" into the frame.
• Requires nodes capable of loading the INT8 ConvRot format; single file is approximately 20 GB; the license does not allow selling; this item has not been tested.

Worth Watching (4)

[B] Tail - Minimax H3 · FL2VA v1.0

LoRA · New Model · Author 2_as_1 · Published 11-04 16:46 · Downloads 88

A concept LoRA for adding a tail to characters in H3 FL2VA videos. The trigger word is tailed.

Reasons for Recommendation

• New Model: The model card body contains only the word "tail". Based on the title and trigger word, the purpose is to add tails to characters.
• 3 visible video examples; free, with a permissive license (selling allowed); 88 downloads (indicating popularity only).
• The same author also released a Qwen Edit version during the same period.

How to Use

• Load tail-minimax_h3-v1.e03000.safetensors (FL2VA version) and write tailed in the prompt; strength and steps are not provided by the author.

Risks and Limitations

• The description is completely blank; tail type, strength, and motion performance must be judged from the examples.
• Another 5 sensitive-level examples are hidden; this item has not been tested.

Trigger Word: tailed

[B] MiniMaxH3. Single frame edit · v0.2

Workflow · Version Update · Author brainburns894 · Published 11-04 14:59 · Downloads 61

A ComfyUI workflow v0.2 that uses H3 as a single-frame image editing model: input a reference image to edit it, supporting multiple references and pose copying. This version can directly output 4K images with optional face refinement.

Reasons for Recommendation

• Version Update: The update note is just one sentence—generates 4K images, optional face refinement.
• Serves image editing: Uses a video model to generate single frames. The author claims it supports multiple references and can copy poses directly.
• The author provides two tips: Clearly describe the style of the input image (e.g., monochrome sketch, anime screenshot) in the detailed_description to preserve the original style; target pixels should not be lower than 2.0 MP.
• Free, with a permissive license; this version has 61 downloads (indicating popularity only).

How to Use

• Drag minimaxh3SingleFrame_v02.json into ComfyUI; specify the input image style in the detailed_description section of the prompt; target pixels ≥ 2.0 MP.
• The required models, nodes, and steps are not specified in the model card by the author; you need to open the workflow to check.

Risks and Limitations

• The model card description is very short, and the dependent nodes and models are not listed; 4 visible examples, another 6 sensitive-level examples are hidden.
• "Preserves 100% original style" is the author's claim; this item has not been tested.

[B] MinimaxH3 swap · v1.0

Workflow · New Model · Author NicknickAI · Published 11-04 13:36 · Downloads 72

H3 video actor replacement workflow: Given a video and several character reference images, replace the people in the video with your characters, keeping camera movement, lighting, background, rhythm, and audio unchanged.

Reasons for Recommendation

• New Model (Workflow): The author claims it only replaces the performer, leaving the original video's camera movement, lighting, background, timing, and sound untouched.
• Serves video character replacement: Insert original characters into existing footage, maintain fixed characters for series content, and use for stand-ins and pre-visualization.
• 1 visible examples; free; 72 downloads (indicating popularity only).

How to Use

• Drag minimaxh3Swap_v10.json into ComfyUI, input the source video and several character reference images.
• The required models, nodes, steps, and number of reference images are not provided by the author.

Risks and Limitations

• The description only covers the purpose, with no parameters or dependency list.
• The license only allows on-site rental, not derivatives; face swapping for real people requires their authorization.
• There are already multiple similar character replacement workflows, and this item has no comparison; this item has not been tested.

[B] KinetiX Weather System Series · Snow & Blizzard v1.0

LoRA · New Model · Author make676767 · Published 11-04 11:09 · Downloads 4 · Paid

A snow weather slider LoRA with positive and negative weights: Positive weights add snow from light snow to blizzards and accumulation, while negative weights melt ice and snow in the scene.

Suitable for creating weather comparisons (-3 / 0 / +3) for the same shot; the author claims future modules will include heavy rain, dense fog, sandstorms, and thunderstorms.

Reasons for Recommendation

• New Model: The snow module of the KinetiX weather series, with clearly defined weight tiers—+1.0~+1.5 for initial snow, +2.5~+3.0 for heavy snow (author recommended), +4.0~+5.0 for blizzards, -2.0~-4.0 for melting snow.
• Serves weather control in text-to-video: Adjusting weights with the same prompt changes snow amount. The author claims buildings and characters do not deform.
• The author provides three directly copyable scene prompts (ancient courtyard, late-night street lamp, snowy pine forest).

How to Use

• No trigger word; add "It is snowing." to the prompt to activate snowfall.
• Set strength_model to +2.5~+3.0 for heavy snow; use -2.0~-4.0 for melting snow.
• Steps and CFG are not provided by the author.

Risks and Limitations

• Paid content; only 1 visible examples (another 3 sensitive-level examples are hidden).
• "Buildings and characters 100% do not deform" is the author's claim; this item has not been tested.

Listed Only (2)

• Lucie Belwyn · MiniMaxH3 — LoRA: Character LoRA for the fictional character Lucie Belwyn, ported from the Anima / Krea 2 version to H3.
• Eminence in Shadow - Minimax H3 - Style · v1.0 — LoRA: H3 style LoRA imitating the art style of the anime The Eminence in Shadow. The author states the effect leans towards 3D and is unsatisfactory.

---

Data: Civitai Public API (Base Model MiniMax H3), collection window 11-04 09:00 → 11-04 21:00 (Beijing Time), backfill period. Recommendation reasons have been reviewed by editors; download counts indicate popularity only.