JiT-DDT

Code: github.com/Linum-AI/jit-ddt  ·  Blog post: Training Text-to-Image Models 3.6x Faster

The same prompt from Linum v2 (2.0B latent-space DiT + VAE, 256x256) and from JiT-DDT (pixel-space, 512x512). JiT-DDT reached this quality in 3.6x fewer GPU-hours and 4.2x fewer samples, at 4x the pixels. The write-up is in the blog post.

This is a research artifact, not a full model release. This model was trained for 138M samples and has not been post-trained. Rather, this is a research preview on the road to our v3 model. We're releasing it because we want to share our preliminary findings with the broader field and encourage others to explore efficient training methods like ours.

JiT-DDT is a 2.5B-parameter pixel-space text-to-image diffusion transformer from Linum: an encoder DiT (64x64-pixel patches, produces a structural plan) and a decoder DiT (32x32-pixel patches, consumes the plan in-context) trained jointly, generating 512x512 RGB images with no VAE. Captions are encoded with Qwen3.5-4B (hidden layers 7/15/27 concatenated).

Code, sampler, reference loss, and the inference code: github.com/Linum-AI/jit-ddt.

Files

File	Contents
model.safetensors	EMA weights, fp32, 598 tensors, 2,487,506,290 parameters (9.95 GB). Keep fp32: the model runs its fp32 master weights under bf16 autocast, which is what the samples in this card were generated with.
config.json	Architecture, text-encoder recipe, sampler defaults, and provenance (source checkpoint step, EMA half-life).
export_manifest.json	Export audit: dropped training-only tensors, key renames, weights SHA256.

Usage

git clone https://github.com/Linum-AI/jit-ddt && cd jit-ddt && pip install -e .
PROMPT="A close-up portrait of a young white woman with vibrant, fiery red hair cascading over \
her shoulders in soft waves, framed from the shoulders up and centered against a softly blurred \
warm-toned background. Her fair, lightly freckled complexion sets off piercing green eyes and a \
subtle, closed-lipped smile. Soft natural light enters from the left of the frame, highlighting \
the texture of her hair and the curve of her cheek while leaving the right side in gentle shadow. \
A shallow depth of field renders the background into smooth, neutral bokeh. Lights dangle out of \
focus on the left side of the frame."
python generate.py --weights Linum-AI/jit-ddt --qwen_model_path Qwen/Qwen3.5-4B \
    --prompt "$PROMPT" --seeds 42

from jit_ddt import JitDDT, QwenTextEncoder, SamplerConfig, generate

model = JitDDT.from_pretrained("Linum-AI/jit-ddt")          # downloads from the Hub
text_encoder = QwenTextEncoder("Qwen/Qwen3.5-4B")
prompt = (
    "A close-up portrait of a young white woman with vibrant, fiery red hair cascading over her "
    "shoulders in soft waves, framed from the shoulders up and centered against a softly blurred "
    "warm-toned background. Her fair, lightly freckled complexion sets off piercing green eyes and "
    "a subtle, closed-lipped smile. Soft natural light enters from the left of the frame, "
    "highlighting the texture of her hair and the curve of her cheek while leaving the right side "
    "in gentle shadow. A shallow depth of field renders the background into smooth, neutral bokeh. "
    "Lights dangle out of focus on the left side of the frame."
)
images = generate(model=model, text_encoder=text_encoder, prompt=prompt, seeds=[42],
                  sampler=SamplerConfig())

The prompt above is the woman_red_hair validation prompt used throughout training and in the blog post.

Sampling: 50 Euler steps, adaptive projected guidance at scale 15, initial noise scale 2.

Provenance

Weights are the EMA (half-life 6,594 steps) trained on 138M image samples in two stages (logit-normal timesteps (0.8, 0.8) then (-0.2, 1.0)).

The 241M-parameter PixelREPA masked transformer adapter (training-only, the representation- alignment loss's projection head) is not included; a reference implementation is in the code repository's loss.py.

Authorship

This model card, and the code repository it points to, were written by Claude (Anthropic's Fable 5.1 model, running in Claude Code) at Linum's request: it extracted the model, inference code and training loss from Linum's internal experiment repository, exported these weights, and verified that they reproduce the training-time validation images bit-for-bit. The model, its training, and the review of this release are Linum's.

License

Apache-2.0. Copyright 2026 Linum Inc.