← sources

High speed speed combat

source: reddit (reddit:StableDiffusion)published 2026-10-27 10:44ingested 2026-10-28 02:37

retained text

Claude code + r2v MINIMAX H3 and some loras + casual prompt like "this shit is not good make it better" My goal is make a 2-5 fucking dinamic anime, but i dont like anime

For anyone interested, here's the current workflow I'm using to make this AI anime.

**Hardware & Setup**

* RunPod with an RTX 4090 (24GB VRAM)
* ComfyUI 0.42.0
* Custom Python automation script (`h3_ctrl.py`)

**1. Character Design & Keyframes**

* **Model:** Krea 2 Turbo (INT8)
* **Text Encoder:** Qwen3VL 4B
* **LoRA:** Kegant V5 (1.0) for that dark, mystical anime aesthetic
* 8–10 steps, CFG 1, Euler/Simple
* Around 12 seconds per image

I generate character reference sheets and keyframes for each scene, including the starting frame and the exact impact frame.

**2. Video Generation**

* **Model:** MiniMax H3 Hybrid FL2VA + REF2VA (INT8)
* **Text Encoder:** Qwen3VL 32B
* **Resolution:** 1344×768, 24 FPS
* **Duration:** 5 seconds per clip, with generated audio

**LoRA stack:**

* Turbo 8-Step (LightX2V): speed and sharpness
* H3 Combat V2: fighting, impacts and dramatic action
* BUNNY Motion Repair V2: smoother continuous movement
* SuperHunman: stronger, more powerful motion
* Action Speed + Acceleration sliders: both at 3.0

**3. The Workflow**

I'm using a modified BUNNY two-pass approach:

* First pass generates the main movement using the LoRA stack.
* Second pass refines details and impacts.
* 10 sampling steps + 4 additional steps for motion formation.
* `res_multistep` sampler.

Character sheets go into `MiniMaxH3ReferenceToVideo`, while `MiniMaxH3AddGuide` anchors the starting frame and impact frame (around frame 72–77).

Prompts are structured into six sections, including detailed character descriptions, continuity constraints, second-by-second action choreography, and sound effects.

**4. Connecting Multiple Clips**

I'm experimenting with **H3 Motion Context by NikoDemon80** to extend scenes beyond 5 seconds.

It takes the last 22 frames of the previous clip as context for the next generation, then removes the overlapping frames when joining everything together.

The goal is longer action sequences with consistent characters and continuous motion.

**5. A Few Important Details**

* No custom-trained character LoRAs.
* I manually check keyframes for costume consistency, enemy counts and character details.
* No artificial speed-ups or slow-motion effects in editing.
* H3 generates the audio too.

And yes, **Claude is involved in the process**, helping orchestrate the workflow and automation.

Still very experimental. Getting consistent, high-speed anime combat is surprisingly difficult, but the results are getting better.

Hope this helps anyone experimenting with H3! 🔥

document fetched 2026-10-28 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xgp5cg/high_speed_speed_combat