akatz-ai/H3-Character-Swap-v1

H3 Character Swap v1

A reference-conditioned character-replacement dataset for MiniMax H3 Ref2VA LoRA training with Ostris AI Toolkit. It combines synthetic still-image edits with unchanged real-motion regularization videos.

134 examples: 94 character-swap edits and 40 preservation clips. Training has 76 edits + 32 clips; validation has 18 edits + 8 clips. Prepared resolution is 1344×768 at 24 fps. The companion 1,000-step LoRA are available separately.

Task and conditioning

An edit example contains:

1. An original scene image, transported to H3 as a five-frame static video: <Video 1>.
2. A character portrait or complete character sheet: <Picture 1>.
3. A short targeting instruction, such as Swap the man in the purple shirt in <Video 1> with the character in <Picture 1>.
4. An edited target PNG containing the replacement character.

The replacement should preserve the reference identity, outfit and reference rendering medium, while preserving the scene's background, non-target people and rendering medium. Lighting, pose, scale and contact adapt to the scene. The 94 edits include 34 cross-style examples and 20 character-sheet examples. Portrait, square and landscape sheets retain their full aspect ratio; sheets remain one image reference. These are one-character swaps, not multi-character-swap targets.

A regularization example uses its same MP4 as both video control and target, with the caption Keep <Video 1> unchanged. Its original soundtrack is retained. No character reference is attached. These are explicit preservation examples. Blank-prompt preservation and preservation with an unrelated character sheet are not included; the package does not claim those behaviors have been taught.

Folder layout

data/
  edits/{train,validation}/
    targets/                CSxxx.png + CSxxx.txt instruction
    character_references/   CSxxx.png, original reference aspect
    scene_videos/           CSxxx.mp4, five static frames at 24 fps
    source_images/          CSxxx.png, original scene at training size
  regularization/{train,validation}/videos/
                            VGxxx.mp4 + VGxxx.txt instruction
metadata/{train,validation}.jsonl
provenance/
  generation.jsonl          generation/edit prompts and original asset hashes
  vidgen.jsonl              original captions, IDs, crop/trim/resize recipes
  curation.json             inclusion and approval record
  upstream/                 snapshot of the source dataset card
configs/
  train.json                full dataset, sampling disabled
  train-with-previews.json   same training recipe, two held-out image previews
  smoke.json                six-step smoke on three edits and three clips
scripts/
  validate.py               records, media, split and checksum validation
  preflight.py              CPU-only installed AI Toolkit compatibility check
  run.sh                    explicit smoke/train launcher
  make_configs.py           regenerate the portable job configurations
  resolve_config.py         write a local UI-ready job with absolute paths
checks/smoke-data/           small copied training-only smoke subset
SHA256SUMS                  canonical package checksums

All paths in the metadata are relative to this directory. Each .txt is the actual short training instruction. Original VidGen descriptions are separate metadata and are never substituted for the preservation instruction. The target PNG remains a single image; repeating the scene source into five frames is conditioning transport, not motion training.

The JSONL tables can be loaded with datasets.load_dataset("json", data_files={"train": "metadata/train.jsonl", "validation": "metadata/validation.jsonl"}). Media columns are path strings, not automatically decoded Image/Video features. Download the complete folder to use the paths. The card follows the Hugging Face dataset-card format.

Media preparation

• Edits: target and scene images 1344×768; native portrait/square/landscape reference aspect preserved. Lossless H.264/yuv444p five-frame scene controls, no audio.
• Regularization: H.264/yuv420p MP4, CRF 18, 24/1 CFR, 73 frames = 17×4+5 = 3.0416667 seconds. Source audio is AAC 192kbps at 48kHz.
• Every regularization source in the downloaded pack was 720p. Selected 1280×720 clips were cropped to 1260×720 (10 pixels each side), then Lanczos-upscaled 1.0666667× to 1344×768. These are not native 768p sources.
• No synthetic detail, optical-flow interpolation, replacement soundtrack or time compression was added. Original source clips remain in the private working collection; only selected normalized excerpts are packaged.
• All 491 matching VidGen captions were reviewed; 82 sources were visually sampled before choosing 40. Final outputs were decoded and frame-sampled. Two clips retain small embedded graphics, and one has softer source detail; see provenance flags.

Preview

Example triplets, ordered source / character reference / edited target:

Source	Reference	Target

Download

hf download akatz-ai/H3-Character-Swap-v1 --type dataset --local-dir ./h3-character-swap-v1

AI Toolkit setup

Use an H3 Ref2VA-capable AI Toolkit runtime. The original checked runtime was an H3-capable AI Toolkit snapshot with a local host-memory compatibility adjustment. Source fingerprints and CPU results are retained in provenance/runtime-contract.json and checks/compatibility-evidence.json. Fresh preflight runs also write checks/preflight-*.json. A stock runtime lacking H3 Ref2VA video-control support is not interchangeable.

The supplied jobs use:

Setting	Value
Architecture	minimax_h3_ref2va, ref2va_pruned
Edit / regularization resolution	1024 area budget → 1344×768 buckets
Edit target / source-video frames	1 / 5
Regularization frames	73 at 24 fps
Edit controls	character reference directory, then scene video directory
Regularization control	exact same directory as targets
Character presentation	image_refs_as_video: false, full-size raw references
Regularization audio	do_audio: true; edits have do_audio: false
Batch / accumulation	1 / 1
LoRA rank / alpha	16 / 16; exclude adaln_proj
Optimizer / pilot learning rate	AdamW8bit / 5e-5
Full pilot steps / saves	1000 / every 250
Compute / quantization	bf16, convrot8 transformer, nvfp4 text encoder
Memory measures	gradient checkpointing, full layer offload, latent/text caches
Guidance loss	enabled, target 3.5
Augmentation / caption dropout	no random crop, flips, or caption dropout

The learning rate and 1000-step budget are starting settings, not a demonstrated optimum. AI Toolkit interleaves the regularization loader with the editing loader; the 76:32 file count is not a 76:32 update weighting. Cache and job outputs belong to the training workspace and should not enter a release.

Install/activate your compatible AI Toolkit environment, then set:

export AI_TOOLKIT_ROOT=/path/to/ai-toolkit
export H3_MODEL_ROOT=/data/models
export H3_OUTPUT_ROOT=/path/to/training-output
python scripts/validate.py --decode
python scripts/preflight.py --toolkit "$AI_TOOLKIT_ROOT" --config train
bash scripts/run.sh smoke
# After inspecting the smoke result:
bash scripts/run.sh train

run.sh sets H3_DATASET_ROOT to this folder automatically. The job JSON files use AI Toolkit's ${ENV_VAR} expansion and can also be passed directly to python run.py with those three H3 environment variables set. JSON is a native AI Toolkit job format. In a UI that does not inherit these variables, use a resolved local copy; do not publish a machine-specific resolved config.

Required model files below H3_MODEL_ROOT:

diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
vae/minimax_h3_video_vae_fp16.safetensors
vae/minimax_h3_audio_vae_fp32.safetensors
loras/minimax_h3_ref2va_training_adapter_v1.safetensors

Weights are not bundled. The smoke job caches only three diverse training edits (including a complex sheet) and three training videos, then runs six steps. It writes to a separate smoke job name. It does not sample images. The default full job also disables automatic sampling; train-with-previews enables two held-out still previews every 250 steps and can need additional memory. Validation examples are never optimization inputs.

Prepared masters versus actual training: configs/train.json describes the original full-resolution proposal; it is not the configuration used for the released checkpoint. configs/trained-run-1000.json records the completed run: edit resolution area budget 1024 (1344×768 buckets), regularization area budget 384 (reduced-resolution video/audio), 73-frame regularization, 1,000 updates, and saves every 250 steps. Sampling was disabled. The trained-run config removes only inactive evaluation sample entries containing local asset paths. The model repository includes the chunked-MLP launcher and dependency snapshot. The runtime was a local source snapshot, so these files are a recorded recipe, not a guarantee of compatibility with every AI Toolkit version. The earlier generic smoke and train configs remain available as proposals and may require more VRAM.

Splits and evaluation

The existing edit split keeps exact source groups and exact reference groups separate; regularization splits keep the VidGen parent video ID separate. Related identities, archetypes or upload channels can still recur, so this is not a certified unseen-identity benchmark. The eight regularization validation clips remain approved for evaluation rather than being folded into training.

Compare the base model and adapter at fixed seeds, references, prompts and sampling settings. Score identity/medium fidelity, correct target selection, unchanged bystanders/background, pose/contact, real-motion consistency, audio preservation, and unwanted swaps. Still-image targets alone do not demonstrate temporally consistent video replacement. Blank-prompt and reference-present no-swap cases need separate evaluation/data.

Provenance and licensing

Synthetic edits were generated with the built-in imagegen tool and represent fictional designs. Generation and correction prompts are retained; CS084 uses the corrected target-v2.png variant. Regularization excerpts come from Fudan-FUXI/VIDGEN-1M, downloaded archive VidGen_video_0.zip; source IDs, original captions and transformations are recorded per clip.

The upstream VidGen card declares Apache-2.0; its exact card snapshot is included. See LICENSES.md, LICENSE, and NOTICE for component licenses and attribution. This package does not label third-party footage as synthetic or public domain. No model weights, credentials, evaluation movie/music/game clips, full million-caption JSON, or unselected videos are included.

This release contains the prepared dataset masters, split metadata, provenance, and preparation/validation scripts. SHA256SUMS covers the release files; runtime caches are excluded. .gitattributes marks PNG/MP4 media for large-file storage; .gitignore lists generated runtime artifacts.