ntc-ai/model-glue-nova-qwen21-experimental

Nova Anime AM ↔ Qwen-Image-2.1: archived bridge experiment

Interactive results · Source snapshot · Checksums

This is an experimental archive, preserved before a new latent-only research direction. It is not a finished general-purpose connector or evidence that independent image models share a lossless latent representation.

The learned inference operation consumes a clean latent estimate and its noise level, and returns a clean latent in the recipient's raw VAE coordinates. It uses frozen pretrained decoder-prefix / encoder-head features plus learned conditional particle-bridge weights. Inference is a deterministic ordinary forward pass, without fitting, retrieval, optimization or best-of-N selection.

Training was not entirely latent-only: target preparation decoded source latents and encoded those images with the recipient VAE, in memory. Training then regressed to cached target latents. No training images were saved or reviewed. This distinction is central to interpreting the archived experiment.

Checkpoints

manifest.json records exact steps, roles, hashes, source revision and stop status. Checkpoint filenames distinguish:

• checkpoints/character-{forward,reverse}.safetensors: the exact 90k / 78k EMA pair used in the published character gallery; these were latest at render start.
• checkpoints/latest-{forward,reverse}.safetensors: latest saved evaluation EMA exports at archive time, distinct from the character snapshot.
• checkpoints/best-{forward,reverse}.safetensors: validation-LPIPS-selected exports at archive time. Character examples did not select these weights.

Only learned bridge weights and fixed calibration buffers are included. Pretrained denoisers, text encoders, VAE weights and optimizer states are not redistributed here. The loader fetches the exact pinned VAE components separately; their upstream licenses apply. This repository's MIT license covers our code and learned bridge artifacts, without replacing upstream terms.

Reproduce inference

Download this repository and unpack source.zip into a separate directory. Install dependencies in a fresh environment; source/README.md describes the pinned numerical stack and includes the original generation commands. The source archive includes the inference, sampler, preparation and training code needed by this experiment, with hashes in source-manifest.json.

import torch
from huggingface_hub import hf_hub_download
from model_glue.image_particle_bridge import load_image_bridge

path = hf_hub_download("ntc-ai/model-glue-nova-qwen21-experimental",
                       "checkpoints/character-forward.safetensors")
bridge, metadata = load_image_bridge(path, "cuda")
# nova_clean_raw: BCHW, [batch, 16, 64, 64] for the evaluated 512px setting.
# Use a finished latent (noise_fraction=0) or an intermediate clean estimate,
# not the noisy sampler state. Output: [batch, 64, 32, 32], Qwen raw VAE units.
with torch.no_grad():
    qwen_clean_raw = bridge(nova_clean_raw, noise_fraction=0.0)

For reverse inference, use character-reverse.safetensors with Qwen raw VAE latents. Shared sampling is a separate operation in model_glue/image_flow_handoff.py; it translates the clean estimate, transports equal-dimensional noise by exact pixel permutation, and resumes the shared Euler trajectory. It does not claim equivalence to each model's native scheduler.

Evidence and limits

The gallery shows two explicitly adult anime characters, both starting models, halfway handoffs and switch-and-return trajectories. Within each row, learned and full-codec controls share prompt, seed, noise grid and denoising-step budget. All eight learned switching trajectories repeated bitwise exactly. Exact execution does not imply exact image fidelity or character identity preservation.

Learned-versus-full-codec LPIPS ranges from 0.0648 to 0.1754 for one switch and 0.1313 to 0.2822 for switch-and-return in these four examples. These are output differences, not aesthetic rankings. The reference trajectory itself often differs substantially from staying in one model.

Validation selects checkpoints; the final test remains unopened. The kingfisher reconstruction is validation, not a final-test example. Character prompts are new development examples excluded from all 4,096 planned training prompts and checkpoint selection. The growing dataset may be incomplete at shutdown; the stop record states actual completed counts. No fixed-cloud control was completed for this warm-started run, so no particle advantage is claimed.

See reports/ for learning histories, sampling measurements, methodology and the final stop record. The GitHub source mirror is 255BITS/model-glue; access to that repository may be restricted. The included source snapshot is publicly available.