akhaliq has concatenated Pika's official Pika-H3-Flex style adapter and MiniMax-H3's few-step distilled LoRA (lightx2v Turbo, pDMD 2 steps) layer by layer into a single standard LoRA: simply drop it into a regular MiniMax-H3 workflow to generate videos with audio in 2, 4, or 8 steps, without needing custom nodes or model modifications. Pika-H3-Flex itself was already included on 10/30, but this combined version was not included in previous issues.
Upvotes 78 · Likes 78 · Downloads 28961
Reasoning
- The author measured (at 848×480, 121 frames with stereo audio, CFG 1): On H200, Pika's official adapter takes about 60 seconds for a 5-second video at 6 steps, while this LoRA takes about 40 seconds at 4 steps and about 25 seconds for a draft at 2 step; on A100, it is 109.3 seconds versus 66.0 seconds. The speedup comes entirely from reducing the number of steps, with per-step time remaining unchanged.
- The method involves rank concatenation, which the author claims is numerically equivalent to loading two LoRAs simultaneously (with relative error per layer on the order of 1e-7), rather than an approximate merge; the Flex style strength is fixed at 0.5 (2 / 4 steps) and 0.6 (8 step HD).
- There are seven files in total, divided into FL2VA (standard text-to-video / image-to-video) and Ref2VA (multi-reference) categories; there is also a lightweight version with rank 32, sized at 0.31 GB, which the author claims retains 96.1% of the spectral energy. The repository includes comparison videos and reproduction scripts.
Usage
- ComfyUI: Select the appropriate file for your workflow and place it in models/loras (use fl2va for standard text-to-video / image-to-video, and ref2va for reference-image workflows), then add LoraLoaderModelOnly to the MiniMax-H3 workflow with a strength of 1.0.
- Sampling: Match the number of steps to the file (2, 4, or 8 steps), set CFG to 1.0, use euler / lcm / res_multistep samplers, and the simple scheduler; use a video shift of 6.0 (FL2VA) or 12.0 (Ref2VA), and an audio shift of 3.0.
- First use the 2 step version to finalize composition and motion, then use the 4 step or 8 step version for the final output; it can also be loaded as a standard LoRA in Diffusers.
Risks
- This is a personal compilation, not an official release from Pika or MiniMax; all speed and visual comparisons were conducted by the author, and the main showcase uses the community merged model Singularity rather than the original H3. This column has not reviewed the comparison videos.
- The author warns: audio quality degrades when using fewer than approximately 6 steps, with the effect being most noticeable at 2 step; the 2 step version occasionally fails entirely and is only suitable for drafting; fast motion is a weakness of few-step models. Do not stack high-strength style or motion LoRAs on top of the 2 / 4 step versions.
- The license follows the MiniMax H3 community license (as a derivative model of MiniMax-H3); check the original terms before commercial use; each file in the standard version is 3.9 GB.