MiniMax H3 Morning Briefing 2026-11-01 | Top Pick: MiniMax H3 Ref2VA + HR Endless — 15 s+ video on a 12 GB GPU

Zhenzhen's AI Radar · MiniMax H3 · 2026-11-01 Morning Briefing

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

Editor's Note This issue reviews 7 new H3 versions, with no S-tier entries. The top recommendation is a long-video workflow for 12 GB GPUs: it splits reference-to-video generation into segments, where Gemma analyzes each segment's visuals to write prompts and continues from the last 22 frames of the previous segment. The author provided measured generation times for 15-second clips, VRAM settings, and known issues like color drift, but only included 1 static frame examples, resulting in an A rating. The H3 version of the retro cyberpunk style LoRA has only two lines of description, and the parameters listed on the card belong to the Z-Image Turbo version, so it was downgraded from A to B. The 2x upscale video VAE in this version only adds int8 format without noting improvements, maintaining a B rating. The three character variants of Game Girl 8 and Renee Hopkins are standard character LoRAs with all examples hidden, rated C.

Recommended (1)

[A] MiniMax H3 Ref2VA + HR Endless — 15 s+ video on a 12 GB GPU · v1 (12 GB tested)

Workflow · New Model · Author semenovsas776 · Published 11-01 03:39 · Downloads 43

An H3 reference-to-video (Ref2VA) long-video workflow: it splits latents into multiple segments generated sequentially. Before each segment, Gemma 4 12B analyzes the previous segment's visuals to write the prompt for the current segment and uses the last 22 frames of the previous segment as video/audio references to continue. The author achieved 15-second final clips on a 12 GB GPU.

Maintained A: Long-video segmented continuation is a clear new use case, with measured generation times, VRAM settings, and known issues clearly documented. However, there are only 1 static frame examples and no final video to view.

Images

Reasons for Recommendation

• New Model: Solves the issue of single-segment H3 videos being too short and hard splicing causing breaks. It serves reference-to-video needs requiring a continuous long shot to tell a story. The [Shot N] storyboard timings written in the prompts are respected.
• The author provided self-test data: On an RTX 4070 SUPER 12 GB, with turbo 4 steps at 480×864, generating 15 seconds / 362 frames took 376–390 seconds. The same workflow with 32 GB RAM took about 425 seconds; 64 GB or more allows the model to stay resident in memory between segments.
• The author compared this with the core node MiniMax H3 Context Windows: The window method is twice as fast (190–204 seconds) and has stable colors, but a single prompt causes looping actions, and one prompt per window causes hard cuts. For a continuous shot with a plot, the author chooses this workflow.
• Free, with a license allowing image output, rental, sale, and sale of fusions. Approximately 5 hours after release, it had 43 downloads (indicating attention only).

How to Use

• Import minimaxH3Ref2vaHREndless15S_v112GBTested.zip (5 KB). Requires ComfyUI-HR-Endless-Sampler (hradec) and ComfyUI-KJNodes. Models are taken from Comfy-Org/MiniMax-H3: ref2va pruned int8, ref2v turbo 4 step LoRA, qwen3vl 32B NVFP4 text encoder, video VAE int8, and audio VAE. The Gemma 4 12B QAT model is automatically downloaded by the node on first run.
• Settings for 12 GB GPUs: chunk_frames 124, video_continuation 22 (full), MiniMax Low VRAM Attention split into 4 head chunks, gemma4_mtp and cache_gemma_preproduction disabled. If later segments run out of VRAM, reduce chunk_frames to 90; if that fails, reduce to 56.
• Actions must fill the entire segment duration (stretching a 6-second action to 15 seconds; Gemma will invent camera moves). At 480×864, distant faces are only about 80–90 pixels high, so place important faces in medium or close-ups.

Risks and Limitations

• Cannot handle dialogue: HR Endless does not support segmented denoising masks, so lines will be voiced by H3 itself. The author measured color drift between segments, with the average saturation in the last second being about 40% higher than in the first second.
• The only visible examples are 1 static frames (a person at a desk); there is no final video. 1 sensitive-level examples are hidden. This entry was not run by us.
• Depends on two custom node packages, five H3 model files, and the Gemma model automatically downloaded on first run. The author's test machine had 96 GB RAM; it will be slower with less memory.

Worth Watching (2)

[B] Cyberpunk 50s · Minimax H3

LoRA · Port · Author Granddyser · Published 11-01 03:15 · Downloads 91

The H3 version of the retro cyberpunk style LoRA "Cyberpunk 50s": transforms visuals into a retro-future atomic age of the 1950s, featuring chrome prosthetics, neon diners, hovercars with tail fins, and rainy boulevards, with a warm Kodachrome film tone.

Downgraded from A to B: The style is not a new capability. The H3 version has only two lines of description and no H3 parameters; the parameters on the card belong to the Z-Image Turbo version. There are only 1 visible examples.

Images

Reasons for Recommendation

• Port: The same model page already has Z-Image Turbo and MOSS Image versions. Shortly after the H3 version was released, a Krea 2 version was also released. The H3 description has only two lines: it can do text-to-video with sound using standard prompts. The author claims the showcase videos did not use trigger words.
• Serves text-to-video needs requiring a unified era art style, such as retro sci-fi shorts, city establishing shots, and character shots. The author claims coverage of characters, street scenes, interiors, and vehicles.
• There are 1 visible example videos (gas station scene: chrome tail-fin car and attendant wearing goggles). Free, with a license allowing image output, rental, and on-site rental. Approximately 6 hours after release, it had 91 downloads, with a cumulative total of 606 for the model (indicating attention only).

How to Use

• Load Retro-Cyberpunk-50s-H3.safetensors (569 MB · FP16) in an H3 workflow and describe the scene using standard prompts. The author did not provide strength, steps, or sampler settings for the H3 version.
• The Shift 7.0, 8 steps, guidance_scale 1, strength 0.7–1.0, and trigger word cy50s listed on the card are under the technical parameters for the Z-Image Turbo version. The author did not specify if these can be directly applied to the H3 version.

Risks and Limitations

• The H3 specific description has only two lines, with no video parameters (frame rate, resolution, steps) and no registered trigger words.
• The author claims showcase videos include full ComfyUI workflows, but the visible 1 segments do not include prompt information. 3 sensitive-level examples are hidden. This entry was not run by us.
• The license does not allow derivatives.

[B] CLEAN X2 E26 VIDEO VAE H3 (2x size upscale) · v2.0

VAE · Version Update · Author speach1sdef178 · Published 11-01 00:45 · Downloads 45

Version 2 of the H3 2x spatial upscale video VAE (Clean X2 E26): It directly outputs double-resolution visuals during decoding, eliminating the need for an external super-resolution model. This version adds int8 files.

Maintained B: The concept is clear and there are video examples, but this update only adds a format without noting improvements or providing comparisons with standard decoding.

Images

Reasons for Recommendation

• Version Update: The update note is just one sentence, "BF16 and Int8 version," meaning int8 files (2.6 GB) were added. The main file is 4.9 GB in fp16 / bf16. The main file hash differs from the previous version, but the author did not state if weights were retrained. The decoding workflow JSON included with the previous version is not attached to this one.
• Serves upscaling after H3 output: The author claims it extracts more spatial detail directly from H3's own decoder, removing the repetitive grid patterns from old X2 decoding, without needing post-hoc AI super-resolution. E26 refers to the 26th major iteration.
• There are 1 visible example videos (2496×1664 close-up of a hen), but no comparison with standard decoding. Free. Approximately 8 hours after release, it had 45 downloads, with a cumulative total of 195 for the model (indicating attention only).

How to Use

• Must first install the TripleHeadedMonkey/ComfyUI-MiniMaxH3_LatentUpscaler node package and use it with the X2 VAE decoding workflow. If VRAM is tight, choose cleanX2E26VIDEOVAEH32x_v20_int8.safetensors (2.6 GB); otherwise, use cleanX2E26VIDEOVAEH32x_v20_fp16.safetensors (4.9 GB).
• The author did not provide decoding chunking, VRAM usage, or recommended input resolution. The X2 decoding workflow is not attached to this version.

Risks and Limitations

• The update note is just one sentence, with no description of visual changes compared to the previous version. The author did not explain the quality difference between int8 and fp16.
• There are only 1 visible examples; 2 sensitive-level examples are hidden. This entry was not run by us.
• The license only allows rental and on-site rental, with commercial image output not checked.

Listed Only (4)

• TFM Game Girl 8 · MMH3 - 3D Version — LoRA: An H3 character LoRA for the author's original mascot character, Game Girl 8 (3D style version): The character originates from a 3D model hand-sculpted by the author, and the training images are multi-angle renders of this 3D model.
• Renee Hopkins · v1.0 H3 — LoRA: An H3 character LoRA for the original fictional adult character Renee Hopkins (a woman living in Halifax who makes her own uniforms to act as an "amateur superhero"), with the trigger word reneehopkins.
• TFM Game Girl 8 · MMH3 3D R2V — LoRA: The H3 "3D R2V" version of the Game Girl 8 character LoRA: It is the same character as the 3D version on the same page, and judging by the version name, it is intended for the reference-to-video (R2V) workflow.
• TFM Game Girl 8 · MMH3 Realistic — LoRA: The H3 realistic version of the Game Girl 8 character LoRA: It places the author's 3D mascot character into H3 videos in a realistic style.

---

Data: Civitai public API (base model MiniMax H3), inclusion window 10-31 20:31 → 11-01 09:01 (Beijing Time). Recommendations have been reviewed by editors; download counts only reflect attention levels.