TheMindExpansionNetwork/swingusinnerz-melt-intermission

Melt Intermission

Eight original-prop surreal cartoon studies commissioned with the source-derived swingusinnerz adapter. Creative direction / LoRA: Mind Expander. Base model: MiniMax H3. Production assistance: Hermes Agent.

Watch

Clip	Intended idea	30-second derivative	~15.083-second source
melt-01	Staircase Soup	silent loop	source with audio
melt-02	Hungry Picture Frame	silent loop	source with audio
melt-03	Accordion Moon	silent loop	source with audio
melt-04	Inside-out Umbrella Universe	silent loop	source with audio
melt-05	Puddle Orchestra	silent loop	source with audio
melt-06	Door Identity Crisis	silent loop	source with audio
melt-07	Self-pouring Teapot	silent loop	source with audio
melt-08	Cricket’s Infinite Curtain Call	silent loop	source with audio

What was actually produced

Exactly eight independently generated 15-second requests, snapping to 362 frames at 24 fps under the supplied H3 frame-grid expression. Public originals preserve the compressed video and generated audio payloads, verified against untouched private pod downloads by SHA-256 stream hashes. Container/workflow metadata is removed. No source/training footage is used as an output clip.

Each derivative is 720 frames / 30 seconds / 24 fps, silent, made by an explicit cosine-eased forward/reverse edit. At output frame j, source frame = round((N-1)(1-cos(2pi*j/720))/2). This traverses the source forward then backward, slowing at turns; it uses nearest-frame resampling, not interpolation. The endpoints deliberately hold near the same source frame. These are edited loops, not a claim that the model generated a perfect seam. Reversed action and paused/duplicated cadence are visible editorial tradeoffs.

Actual observed descriptions and limitations are in manifest.json and prompts.json; titles name the intended ideas, not promises that every transformation rendered. QA sheets in qa/ show source progression, loop samples, decoded boundaries and mid-cycle reversals. Full decoding was checked for every private original, public original and loop. Repeated playback was reviewed separately; a low boundary error is supporting technical evidence only.

Inference and provenance

The supplied VDN graph is preserved: FL2VA pruned INT8 ConvRot base, stage-dmd-step-250, unchanged turbo adapter, 8 Euler/simple steps, 0.8 MP, 24 fps. Only the five authorized leaves in the manifest changed for generation. One exact swingusinnerz_t2v_minimax_h3_v1.safetensors scene adapter at strength 1; no stacking of scene LoRAs. Sanitized workflows retain the execution graph and replace output prefixes only.

Pinned model config: https://huggingface.co/TheMindExpansionNetwork/swingusinnerz_t2v_minimax_h3_v1/blob/f1ce7f1183b1d156f84e469d18e4080f840042b2/config.yaml

Source dataset reference: https://huggingface.co/datasets/TheMindExpansionNetwork/swingusinnerz-caption_qwen3vl_omni

The config's trigger_word is null. Literal swingusinnerz, comes from its sample prompt, not a proven special token. model-config-sanitized.yaml redacts local filesystem paths; manifest records the original upstream config SHA-256 and revision. The installed adapter filename was verified, but installed weight bytes were not independently matched against Hub weights. Training-preview visual reference showed monochrome elastic cartoon silhouettes; prompts request original props and a new cricket, not historical characters or copied source shots. No reference images or private training captions are redistributed here.

Disclosures and rights

• AI-generated source-derived rubber-hose experiments; intended prompts are not output guarantees.
• Exactly eight new source renders; requested 15 seconds, actual 362 frames at 24 fps (about 15.083 seconds). No extra rerenders.
• Loops are 30-second silent editorial cosine-eased forward/reverse cycles. Backward motion, nearest-frame sampling and slow/held turns are intentional. No claim of naturally generated seamlessness.
• Public originals are metadata-stripped remuxes preserving compressed generated video and audio payloads; untouched files remain private.
• Source audio is preserved, not transcribed or rights-cleared. Visual-only loop audio is intentionally absent.
• SWINGUSINNERZ is a source-derived adapter related to Swing You Sinners! Historical characters, film transfers, restoration, soundtrack, LoRA, base-model and output rights are separate. No worldwide clearance or new blanket license is asserted. Review applicable MiniMax H3 terms and local law before reuse.
• Visual QA uses actual contact sheets, decoded boundary metrics and repeated-motion review; smooth endpoints do not eliminate internal generative morphing or guarantee subjective seamlessness.

LoRA/style provenance relates to Swing You Sinners!; an original prompt does not itself establish output non-infringement. Public availability of this dataset is not a license grant for third-party material. No broad source, audio, base-model or generated-output clearance is claimed. No license field is assigned on this card. Review the current MiniMax H3 license at its official model repository and any adapter/source terms separately.

Files and download

• originals/: privacy remuxes, generated audio retained.
• loops/: silent edited derivatives, H.264 yuv420p / faststart.
• posters/: JPEG frames extracted from actual originals.
• workflows/, model-config-sanitized.yaml, manifest.json, prompts.json: reproducibility, intended vs observed notes and provenance.
• qa/: actual contact sheets, boundary and reversal sheets.
• SHA256SUMS: every other file's byte hash.

hf download TheMindExpansionNetwork/swingusinnerz-melt-intermission --repo-type dataset --local-dir ./melt-intermission
cd melt-intermission
sha256sum -c SHA256SUMS

For immutable reproduction, add --revision <verified commit> from the publication receipt. The parent festival maintains the screening link separately; this pack does not modify the festival.