═══════════════════════════════════════════════════════════════════════════════ H3 TURBO 4-8 STEP (ADALN-PRUNED) ═══════════════════════════════════════════════════════════════════════════════ ─────────────────────────────────────────────────────────────────────────────── IDENTITY ─────────────────────────────────────────────────────────────────────────────── Base model MiniMax H3 Generation mode First-and-Last-Frame-to-Video (FL2VA), Reference-to-Video (Ref2VA) Checkpoint both families — see the author's builds Category Speed / step-reduction Author Rkss Source https://huggingface.co/Rkss/Minimax_h3_fl2v_lightx2v_turbo_4to8step_v0.1-v1.0_768p_v4_step600_dareties_fro095_native Provenance HuggingFace · Rkss/Minimax_h3_fl2v_lightx2v_turbo_4to8step_v0.1-v1.0_768p_v4_step600_dareties_fro095_native · minimax_h3_fl2v_lightx2v_turbo_4to8step_v0.1-v1.0_768p_v4_step600_dareties_fro095_pruned.safetensors ─────────────────────────────────────────────────────────────────────────────── WEIGHTS ─────────────────────────────────────────────────────────────────────────────── File minimax_h3_fl2v_lightx2v_turbo_4to8step_v0.1-v1.0_768p_v4_step600_dareties_fro095_pruned.safetensors Version main Size 818.9 MB Format SafeTensor / unknown precision SHA256 063df5592549161ff62a1e29ef49978625ee5060eda7d97edfa631bd9aa713b1 Published 2026-09-04 ─────────────────────────────────────────────────────────────────────────────── TRIGGER WORDS ─────────────────────────────────────────────────────────────────────────────── No trigger word. The LoRA applies as soon as it is loaded — strength is the only way to dial the effect up or down. ─────────────────────────────────────────────────────────────────────────────── RECOMMENDED SETTINGS ─────────────────────────────────────────────────────────────────────────────── Taken verbatim from the author's description: Sampler steps 4–8 ─────────────────────────────────────────────────────────────────────────────── ADOPTION & POPULARITY ─────────────────────────────────────────────────────────────────────────────── Downloads 941 Rating 100% positive (1/1) Comments 0 Published 0 days ago Download rate 941.0/day since release Adoption rank #38 of 113 in this collection (Moderate) Momentum rank #12 of 113 by download rate (Hot · fresh release) Adoption tier is the percentile of total downloads within this collection; momentum is the percentile of downloads-per-day since release. Momentum is ranked rather than measured against a fixed rate, because this base model is itself new and every LoRA looks fast on an absolute scale. Both numbers are CivitAI's, read on the scrape date at the foot of this file. ─────────────────────────────────────────────────────────────────────────────── CHAINING / STACKING ─────────────────────────────────────────────────────────────────────────────── Stacks with others not stated The author gave no chaining guidance. The rules below are the general MiniMax H3 ones, not anything specific to this LoRA. General MiniMax H3 rules that apply here (see REFERENCE.txt): • Load the LoRA with LoraLoaderModelOnly between UNETLoader and your sampler; stacking means chaining several of those nodes in series. • This is a step-distillation (turbo) LoRA. DISTILL LORAS DO NOT COMPOSE — running this alongside another turbo/PDD/lightx2v LoRA breaks generation. Pick exactly one. • Style and character LoRAs stack on top of it normally, with no special strength penalty beyond ordinary multi-LoRA tuning. • Use Euler so the trained sigma schedule is consumed as intended, and keep CFG at 1.0 with BasicGuider. • Only the step counts the LoRA was trained for (usually 4, 6 or 8) work. Other values produce noise — it is not a free dial. • The author ships builds for both checkpoint families — pair each build with its own checkpoint and never cross-pair them. ─────────────────────────────────────────────────────────────────────────────── PROS ─────────────────────────────────────────────────────────────────────────────── + Author documents use across several modes (FL2V, REF2V). [source: description] + Author published concrete recommended settings. [source: description] ─────────────────────────────────────────────────────────────────────────────── CONS & CAVEATS ─────────────────────────────────────────────────────────────────────────────── - No trigger word — the LoRA is always on once loaded, so strength is the only control. [source: model metadata] - Large at 819 MB — slower to load and heavier to stack. [source: file size] - Modest adoption (941 downloads) — less community feedback to rely on. [source: source stats] ─────────────────────────────────────────────────────────────────────────────── COMPATIBILITY WARNING ─────────────────────────────────────────────────────────────────────────────── MiniMax H3 ships two transformers that share tensor key names: the fl2va checkpoint (driving T2VA, I2VA, L2VA and FL2VA) and the separate ref2va checkpoint (driving Ref2VA). Loading a LoRA against the wrong family can silently degrade output instead of raising an error, so confirm the family above matches the checkpoint you have loaded. Likewise, LoRAs built for full checkpoints often throw adaln_proj shape errors on pruned ones — match the build to your checkpoint. ─────────────────────────────────────────────────────────────────────────────── AUTHOR'S DESCRIPTION ─────────────────────────────────────────────────────────────────────────────── The second of this repo's two builds of the same Turbo step-distillation LoRA: this one has the AdaLN modules pruned (624 tensors vs. 726 in the native port; its metadata states the only intentional information removal is all 51 AdaLN LoRA pairs, with no SVD or requantization of the retained factors). Same 4-8 step acceleration and strength guidance as the native build. Distinct weights, not a re-upload. ─────────────────────────────────────────────────────────────────────────────── EXAMPLE MEDIA ─────────────────────────────────────────────────────────────────────────────── 1 preview(s) in ./example_images/ — community results for this LoRA. Videos keep a matching _poster.jpg still frame. No prompts are recorded below: CivitAI's public API flags these posts as having generation metadata but does not return it, so the prompts cannot be scraped. Open the model page to read them. 01_image.png (image, ) ─────────────────────────────────────────────────────────────────────────────── HOW TO USE (COMFYUI) ─────────────────────────────────────────────────────────────────────────────── 1. Copy the .safetensors file into ComfyUI/models/loras/ 2. Load the matching H3 checkpoint with UNETLoader. 3. Insert LoraLoaderModelOnly between the loader and your sampler. 4. Set strength as listed above, then add any trigger word to the prompt. 5. Keep CFG at 1.0 with BasicGuider unless you know you want real CFG. ─────────────────────────────────────────────────────────────────────────────── Scraped by lorakit on 2026-09-05 01:13 UTC. Stats and description are the author's; pros/cons are derived from the evidence tagged beside each line. ───────────────────────────────────────────────────────────────────────────────