NextStep-1¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
nextstep-1-large |
stepfun-ai/NextStep-1-Large | 97dd0b9 |
Apache-2.0 | Verified |
Catalog: nextstep-1 ·
Model: nextstep-1-large · Task: t2i
NextStep-1 Large¶
Checkpoint. nextstep-1-large pins the author repository
stepfun-ai/NextStep-1-Large to revision
97dd0b9e097f60e39e3ce3bbac1a3343a09547cd. The architecture reference is the official
NextStep-1 repository at a3303b62e11fc7cd473ba0e4b1f32eda73f4acc0.
Architecture. NextStep-1 combines a 48-layer, width-5120 Qwen2.5-derived causal decoder with continuous image tokens. A learned input projector embeds each packed 2×2 latent patch. The final hidden state conditions a 12-layer adaptive-normalization MLP that integrates one new latent patch with flow matching, after which the patch is fed back into the decoder. A 16-channel KL autoencoder decodes the completed latent grid.
Generation. Prompts use the published image-area suffix and beginning-of-image token. Image patches are generated left to right with a persistent key/value cache. Guidance uses positive and negative prompt branches; each patch is sampled by the published ODE or SDE flow update, unpacked, and decoded to uint8 pixels.
Settings. Defaults are 256×256, 20 flow steps, guidance 1, timestep shift 1 and SDE sampling.
width and height must be positive multiples of 16. The backend also accepts seed,
negative_prompt, steps, guidance, timestep_shift and sde. Unknown settings are rejected
before loading.
Validation and limits. Offline tests check every published transformer tensor name and shape, exact tiny random-weight decoder and flow-head parity, option validation and deterministic tiny generation. Separate opt-in end-to-end cases compare seed 0 and seed 42 through the public pipeline at 64×64 with one flow step. Those deliberately short cases retain prompt formatting, causal cache, flow integration, latent unpacking and VAE decoding without running the very large checkpoint in the fast suite. Real-checkpoint tests are provided but are not run as part of this implementation step. Batches, image editing, training and non-float32 parity are out of scope.
Memory. The twelve bfloat16 transformer shards total 58.3 GB (about 54.3 GiB), plus a 335 MB VAE checkpoint. Loading the weights upcast to CPU float32 requires substantially more memory; the reference end-to-end case therefore runs in a separate process and exits before native loading.
License. The checkpoint and official code are Apache-2.0. imagehub does not distribute weights.