DanichAI — fixed-grid token-video research

Research update: 9 October 2026

DanichAI is being rebuilt as a custom video architecture. The current path does not use Wan diffusion, MaskGIT, Gaussian image noise, or an external image decoder. A frame is a fixed 704×704 field of addressable tokens, initially white, and the model changes the semantic parameters of those tokens.

This page reports measured intermediate results. It does not claim that the final prompt-to-video generator is complete.

Token representation

Every token has a fixed [y, x] address and four semantic parameters:

1. brightness: integer 0..100;
2. base colour: integer 0..14;
3. blur/sharpness: integer 0..100, where 0 is sharp and 100 is blurry;
4. sparse exact RGB correction: one vector parameter with three uint8 channels.

The exact correction uses a dead zone and regularization so it does not simply become a hidden RGB copy. On excluded codec data, 32.72% of correction channels are exactly zero.

Video memory is addressed as [frame, y, x, token parameter]. The implemented 1,080-frame memory has shape [1080, 704, 704, 6]; six stored channels represent four semantic parameters because the RGB correction is a three-component vector.

Motion update policy

The temporal design preserves the previous frame and updates selected regions:

• red: must change;
• yellow neighbors: may change to maintain boundaries and motion;
• green background: remains cached unless a distant update is requested.

The controller may select disconnected regions. The first frame is drawn from a white field; subsequent frames should update motion and neighboring tokens.

Custom codec holdout

Three source videos were completely excluded from training. Results below use 12 full 704×704 frames from those sources:

These measurements validate the custom encoder/decoder, not an independent scene generator.

Spatial layout model

The accepted model keeps a 176×176 spatial latent instead of the earlier 88×88 grid. It was trained on 51 sources while the three evaluation sources remained excluded.

V7 improves every listed metric and passed the same immutable holdout gate.

Independent scene-prior pilot

After accepting v7, training moved from codec refinement to a separate scene prior. Its input is a 128-bit discrete binary scene code rather than Gaussian noise. The generator begins from the encoded white fixed-grid field and predicts pairs of consecutive 176×176 layout latents. A pair discriminator learns the distribution and temporal relation of real adjacent frames from 51 sources.

The first two scene-prior checkpoints were rejected after rendering unseen binary codes. Neither copied a checked source frame, but both failed temporal coherence:

V3 adds layout-manifold cycling, real-motion matching, and spatial-gradient matching. In parallel, an autonomous search evaluates 216 combinations of binary-code size, model width, learning rate, and loss profile. Static collapse is explicitly penalized; 12 candidates receive longer training and three finalists receive full 96-frame audits. Training activity alone is not treated as a quality result.

Measured generation and 4K output speed

The current layout composition pilot at 704×704 reaches 160.37 frames/s with batch 16 on an RTX 5090 and produced no exact source-frame matches in the checked run. It is still a learned composition of input sequences and is not proof of independent scene generation.

Three parallel CUDA-scaling/NVENC sessions converted a 12 FPS unique-frame stream into 24 FPS CFR 4K video:

These are component measurements. A concurrent one-hour generator-to-4K run is still required before the five-minute end-to-end target can be claimed.

Victory gate

DanichAI will be called successful only after all of these pass:

• a new coherent video generated from the model's own scene state;
• stable localized motion with cached background;
• no exact copying and a nearest-neighbor audit;
• one complete concurrent 4K/24 FPS throughput run;
• released matching code, weights, and reproducible holdout metrics.

Scene-by-scene prompt understanding is deliberately deferred until this base video gate passes. The planned scene representation includes persistent object IDs, actions, camera, lighting, duration, transitions, and token-region maps.

Current fixed-grid component hashes

The accepted fixed-grid codec and spatial-layout checkpoints are downloadable from this repository. Their hashes allow each download to be checked byte for byte:

weights/danich_four_parameter_codec_diverse_v5.pt
weights/danich_layout_ae_4x_v7.pt

SHA-256:

danich_four_parameter_codec_diverse_v5.pt
5872826EC2087A0928BAF517EBB0041CD05AEEA2D40A925793596F465B213A8E

danich_layout_ae_4x_v7.pt
6C61C06503EEB773C5226E795F1DEFE3472D3986F4ABF0527857E6CB8CE0802E

These are validated components, not a completed prompt-to-video generator. Scene-prior weights remain unreleased until the coherence gate passes.

Legacy 4 October checkpoint

The older Wan-latent decoder remains available for reproducibility:

weights/danichai_decoder_wan_sharp_release_20261004.pt
src/vqvid.py
src/vq_token_train.py
src/lfq.py

SHA-256:

03157266249FFF9F42C4B3EAB48FAE1E122E969255E9C27F55D86EFA6BC3A396

This hash is verified against the uploaded legacy file. It is not the new fixed-grid codec described above.

Limitations

• The independent white-field scene prior is not yet validated.
• Current novel-video tests remain compositions of learned source sequences.
• The 4K result validates the output stage, not a complete concurrent hour run.
• Native model resolution is 704×704; 4K currently uses GPU scaling.
• Audio is not implemented.

License

Code and weights are licensed under AGPL-3.0-or-later. Network use requires offering the complete corresponding source to users of the service. Third-party components retain their own notices.

Responsible use

DanichAI can synthesize video, including people and faces. Do not use it for impersonation, fabricated evidence, or non-consensual deceptive media.