STUUBZZZ
MiniMax H3 timeline in ComfyUI: middle frames, extending a clip, pinned sounds and clips
Ultimate H3 0.11 lets you direct MiniMax H3 in ComfyUI with up to two middle frames, a clip continued from its last 1.6 s, and up to 8 pinned sounds or clips.
Blog · 19 posts
ComfyUI notes from the pipeline
Settings, measured timings and teardowns for self-hosted AI video in ComfyUI: MiniMax H3, LTX 2.5, the Stubelius workflows, LoRA training and character consistency. My own timings come from one RTX 5090, and each post says how it was tested and where a number is someone else's.
All posts
ComfyUI errors beginners hit: what each of the eight means and how to fix it
The free Part 1 of ComfyUI From Zero covers eight errors beginners actually see, from Value not in list to CUDA out of memory: each message, meaning and fix.
MiniMax H3 in 4 steps: PDMD LoRA vs NVIDIA's SoL Refiner, timed on an RTX 5090
PDMD's 4-step LoRA finished a 6.6 s MiniMax H3 clip to 2520×1440 in 101 s on an RTX 5090, against 753 s for 20 steps plus polish. SoL Refiner changed faces.
LTX 2.5 modes in ComfyUI: what each one does and how to stop MSR duplicate characters
Each LTX 2.5 mode in ComfyUI wants its own prompt. With MSR V1, naming @ref1 in every action cut seeds with a duplicate character from 2 of 4 to 0 of 4.
MiniMax H3 ComfyUI workflow: Speed, Hybrid and Quality modes timed on an RTX 5090
Stubelius Ultimate H3, a free MiniMax H3 workflow for ComfyUI: three presets, 8 s clip times and peak VRAM on an RTX 5090, seed hunting and install.
LTX 2.5 blurry in ComfyUI: render the first pass at full size
Soft LTX 2.5 video often comes from a half-size first pass. Same seed on an RTX 5090: 61.1 s at full size vs 46.2 s at 0.5 plus a x2 latent refine.
MiniMax H3 sampler and scheduler settings: three safe pairs, two that break
Five MiniMax H3 workflows at 6 to 14 steps: euler/simple, euler/beta and er_sde/beta had no gross failure. dpmpp_2m_sde/karras gave noise on four of five.
LTX 2.5 VAE decode: flicker every 56 frames and the shared-memory stall on Windows
LTX 2.5 flickers every 56 frames when VAE Decode (Tiled) uses temporal overlap 8. On Windows an oversized decode spills into shared GPU memory. Two fixes.
DLSS 5 vs RTX Video Super Resolution in ComfyUI: upscale times and keeping the colours
Upscaling an 8 s MiniMax H3 clip to 1440p in ComfyUI on an RTX 5090: RTX VSR 1:06, DLSS 5 + Color Lock 4:27. Plus LTX 2.5 times and the colour fix.
MiniMax H3 prompt format: three fields, speaker tags and a word budget
The structured prompt MiniMax H3 expects: alignment line, three named fields, (S1) speaker IDs, dialogue tags, shot timestamps and the 17k+5 frame grid.
MiniMax H3 TaoMate LoRA, Singularity and turbo: which pairings hold colour
Eight MiniMax H3 runs on one seed: TaoMate 3-step on Singularity v1.3 drops saturation from 42.8 to 28.6. TaoMate on stock and turbo on both hold colour.
AI Toolkit MiniMax H3 LoRA training: what the Distillation Handling Method setting does
From the AI Toolkit code at commit ecee894: what Training Adapter and Contrastive Guidance set for MiniMax H3 LoRAs, and when Differential Guidance runs.
Stubelius Ultimate LTX 2.5 ComfyUI workflow: multi-shot timeline, seed hunt, Licon MSR
Free LTX 2.5 ComfyUI workflow: multi-shot timeline, 1 to 4 seeds with sound, Licon MSR references, up to 4K. Install steps and RTX 5090 timings.
MiniMax H3 Director in ComfyUI: seed hunt, two-stage sampling and the Stubelius refine
Stubelius Director for MiniMax H3 in ComfyUI: scout up to 4 seeds at 0.25 MP, finish one with a learned 2x latent upscale. Settings and install.
FastH3 vs Turbo LoRA on MiniMax H3: 5.1 vs 6.21 minutes per clip on an RTX 5090
FastH3 preview adapter at 6 steps vs the 8-step Turbo LoRA on MiniMax H3: 5.1 vs 6.21 min for a 15 s clip on an RTX 5090, same time per step.
FastH3 LoRA does nothing in ComfyUI: why, and the converter that fixes it
FastVideo's FastH3 preview adapter loads in ComfyUI and changes nothing: 0 of 389 module names match the pruned MiniMax H3. Key map, converter, settings.
LTX 2.5 sampler settings: euler_ancestral or res_2s, and when you need the distilled LoRA
LTX 2.5 distilled in ComfyUI: euler_ancestral + linear_quadratic at 8 steps, cfg 1, matches Lightricks' sigmas. Plus when the distilled LoRA is needed.
MiniMax H3 2.5D anime style LoRA: what the retrain changed and which strength to use
MiniMax H3 slides to flat anime or 3D. A style LoRA retrained on 50 three-second clips holds a glossy 2.5D look. Same-seed strength sweep from 0.4 to 1.0.
Consistent character LoRA dataset: the five views and the rules
The 15 to 30 images for a character consistency LoRA: front, side, 3/4, back and face detail, what to lock, how to caption, and what changes for video.
Want the why, not just the workflow?
Learn it, fix it live, or have it made.
The ComfyUI From Zero course explains the machine from the first node to training your own LoRA, and Part 1 is free. If something is fighting you right now, bring it to a 60-minute 1-on-1.