TheMindExpansionNetwork/swingusinnerz_t2v_minimax_h3_v1

SWINGUSINNERZ — MiniMax H3 T2V LoRA v1

An experimental MiniMax H3 text-to-video LoRA for the SWINGUSINNERZ source-study/channel project: monochrome animation, elastic motion, surreal transformations, theatrical staging, and audio-visual cartoon rhythm.

Status: trained adapter and training-preview videos are uploaded. Fast-mode inference, realtime throughput, deployment compatibility, and generated-audio quality have not been independently benchmarked for this release. This is an adapter, not a standalone base model or an operational live channel backend.

Project links

• SWINGUSINNERZ channel / source-reference player
• Public 15-second source clips with audio
• Improved caption dataset
• Channel library
• Show Engine prompt builder
• Private channel control project — access required

These are related project resources. The training config records a local dataset folder, not a Hub dataset revision or file-hash manifest; exact correspondence between the uploaded training run and a particular caption-dataset revision still needs a training receipt.

Download

hf download TheMindExpansionNetwork/swingusinnerz_t2v_minimax_h3_v1 \
  swingusinnerz_t2v_minimax_h3_v1.safetensors \
  config.yaml README.md \
  --local-dir ./swingusinnerz-h3-v1

Download only the adapter and documentation for inference. optimizer.pt, loss_log.db, log.txt, and pid.txt are training artifacts, not inference dependencies. Do not load an untrusted pickle-based optimizer checkpoint just to use a LoRA.

What the uploaded configuration says

Source: config.yaml, cross-readable with .job_config.json. The configuration was inspected at repository revision 0b24e01544089131038c80672227d407a17ce716.

Setting	Recorded value
Base model	Comfy-Org/MiniMax-H3
Trainer process / architecture	diffusion_trainer / minimax_h3
Network	LoRA
Linear rank / alpha	16 / 16
Convolution rank / alpha	16 / 16
Ignored module-name fragment	adaln_proj
Configured training steps	3000
Batch size / gradient accumulation	1 / 1
Learning rate	0.0001
Optimizer / weight decay	adamw8bit / 0.0001
Scheduler / loss	flowmatch / mse
Training and save dtype	bf16
Text encoder training	Disabled
Gradient checkpointing	Enabled
Caption dropout	0.05
Dataset resolution setting	[512]
Dataset frame setting	39, with auto_frame_count: true and shrink_video_to_frames: true
Dataset FPS	24
Image-to-video dataset mode	do_i2v: false
Audio dataset mode	do_audio: true
Audio loss multiplier	1
Model quantization	convrot8
Text-encoder quantization	nvfp4
Training assistant adapter	ostris/minimax_h3_training_adapter/minimax_h3_training_adapter_v1.safetensors

The config also enables guidance loss (guidance_loss_target: 3.5) and differential guidance (differential_guidance_scale: 3). These are training settings, not suggested inference CFG values. The assistant adapter path describes training setup; it does not by itself establish that the assistant adapter must be stacked at inference.

Configured values describe the job recipe, not independently measured hardware use, loss quality, exact effective frame counts, or successful runtime support on another installation.

Training previews

The uploaded sampling recipe is:

• 512 × 512
• 73 frames at 24 FPS
• FlowMatch sampler, 28 sample steps
• Guidance scale 1
• Seed 42, with walk_seed: true
• A preview every 250 steps

Watch the step-3000 training preview

Browse all training previews

The repository also includes intermediate adapter files at steps 2000, 2250, 2500, and 2750, alongside the unsuffixed release adapter. A later training step is not automatically a better checkpoint; compare motion, artifacts, prompt adherence, and audio before selecting a channel preset. Preview comparisons are not necessarily fixed-seed comparisons because seed walking is enabled.

Aspect-ratio note: the stored sample prompt describes a 4:3 frame, but the actual configured preview dimensions are square. Prompt text does not override width/height settings.

Prompt namespace

The training config has trigger_word: null. Its sample prompt begins with the literal text swingusinnerz,. Therefore this README recommends that spelling as a starting prompt prefix, not as a guaranteed learned special token.

The related caption project has used <swingusinnerz> as a namespace marker. The config alone does not prove that every training caption used that bracketed form. Preserve actual training-caption spellings when comparing results; do not assume the two forms are equivalent.

Original-scene test prompt

swingusinnerz, integrated_multimodal_description: An original monochrome animated scene on a small theatrical stage. A crooked umbrella grows rubbery legs, tiptoes around a puddle, and folds itself into a bow. Elastic motion, hand-drawn line variation, restrained film grain, and one clear transformation. The camera holds a wide composition. overall_soundscape: Soft rhythmic footsteps, a springy folding sound, and a gentle splash synchronized to visible actions. non_diegetic_music: An original playful instrumental rhythm, with no borrowed melody or dialogue.

This is an untested example, not an output-quality guarantee. Start with one subject, one action, and one transformation. Compare LoRA-off and LoRA-on renders with otherwise matching settings. Do not assume a universal LoRA strength, accelerated step count, or loader format: those depend on the H3 implementation and workflow.

Fast-mode / channel integration

The next experiment is a generated SWINGUSINNERZ playback lane, separate from its archival reference player.

1. Load the compatible H3 base and this adapter in a verified H3-capable workflow.
2. Establish a baseline using the recorded sampling recipe where supported.
3. Test fast-mode or acceleration adapters separately; record their exact versions, order, strengths, and scheduler settings.
4. Check actual video dimensions, frames, FPS, audio-stream presence, audiovisual alignment, and prompt adherence.
5. Record wall-clock generation time and hardware before making throughput claims.
6. Queue reviewed clips into a buffered player with a prerecorded fallback.

No realtime or particular GPU/VRAM performance is promised. Audio-enabled training does not prove every inference workflow will emit audio. The related 15-second source-pack delivery format is not evidence that this run trained on full 15-second clips; its dataset and preview settings above differ. Longer renders require separate testing.

Provenance, rights, and limitations

This is a source-derived experimental adapter, not an ownership claim over the historical film Swing You Sinners! or its soundtrack. The related source datasets carry provenance and U.S.-first rights notes. Consult those records for the exact source transfer and captions.

• Public-domain conclusions are source- and jurisdiction-specific; they are not blanket worldwide clearance.
• Audio recordings, scores, restorations, transfers, and other added material may have separate rights or uncertainty.
• Base-model terms, adapter permissions, source rights, and output use are separate questions. No new adapter license is assigned by this README.
• Training can reproduce source details. Review outputs for copied characters, shots, compositions, melodies, and dialogue before release; prefer original scenes and subjects.
• Namespace labels are not ownership tokens. Do not label archival/source-derived material as <UL_ORIGINAL_V1>.
• Present generated channel material as AI-GENERATED and archival examples as SOURCE REFERENCE.

Reproducibility gaps

Still needed: exact training-data revision and hashes, trainer commit/version, final checkpoint-step metadata verification, inference workflow export, tested LoRA strength, hardware/VRAM measurements, and independent audiovisual evaluation. The presence of uploaded weights and previews is not a substitute for these checks.