Skip to content
Bud's Wiki

Nucleus-Image — sparse MoE diffusion model, 17B total / 2B active (Nucleus AI announcement)

TL;DR

Section titled “TL;DR”

Nucleus AI announces Nucleus-Image, claimed as the first sparse Mixture-of-Experts diffusion model: 17B total parameters with only 2B active per forward pass, positioned as ~10× more parameter-efficient than dense leading diffusion models. The base model (pre-training only, no DPO / no RL / no preference tuning) is claimed to be competitive with GPT Image 1, Imagen 4, and Qwen-Image on standard image-generation benchmarks. Release is Apache 2.0 with weights, training code, and dataset recipe, plus day-0 Hugging Face diffusers support. A second tweet in the thread reports the headline benchmarks: DPG-Bench 88.79 (#1 overall), GenEval 0.87 (matching Qwen-Image, beating GPT Image 1), OneIG-Bench 0.522 (beating Imagen 4 and Recraft V3).

Key claims

Section titled “Key claims”
  • Nucleus-Image is positioned as the first sparse MoE diffusion model, with 17B total parameters and 2B active per forward pass [tweet 1].
  • Claimed parameter efficiency: ~10× fewer active parameters than dense “leading diffusion models” for comparable quality [tweet 1].
  • Base-model performance (no DPO, no RL, no preference tuning) is claimed competitive with GPT Image 1, Imagen 4, and Qwen-Image [tweet 1].
  • Reported benchmarks (base model, no post-training): DPG-Bench 88.79 (“#1 overall”), GenEval 0.87 (matches Qwen-Image, beats GPT Image 1), OneIG-Bench 0.522 (beats Imagen 4 and Recraft V3) [tweet 2].
  • The accompanying scatter plot positions Nucleus-Image at 0.76 “overall performance” in the low-active-parameter / high-quality quadrant, with 2B active parameters vs Qwen-Image at 20B active for similar overall score [tweet 2, plot].
  • Release package: Apache 2.0; weights, training code, and dataset recipe; day-0 Hugging Face diffusers integration [tweet 1].

Method

Section titled “Method”

Tweet content only — no technical report or arXiv companion linked at announcement time. The architectural details (MoE granularity, expert count, routing mechanism, base backbone, training recipe) are not disclosed in the thread. What is stated: a sparse MoE structure with a 17B parameter pool and 2B active per step, trained purely with pre-training (explicitly no preference data, no RL stage, no DPO). The accompanying benchmark scatter compares against GPT Image 1 High, Qwen-Image, Seedream 3.0, FLUX.1 Dev, HiDream, CogView 4, and others on an unspecified composite “overall performance” metric.

Results

Section titled “Results”

The headline numbers reported in the follow-up tweet, base-model only:

  • DPG-Bench: 88.79, claimed #1 overall.
  • GenEval: 0.87 — claimed to match Qwen-Image and beat GPT Image 1.
  • OneIG-Bench: 0.522 — claimed to beat Imagen 4 and Recraft V3.
  • Composite “overall”: ~0.76 at 2B active parameters, placed in the top-left “low cost, high quality” region of the comparison scatter.

No FID, no human-preference winrate, no ablations, no comparison to MoE diffusion variants. The 49.9K-view first tweet and 2,989-view second tweet are the only public surface at filing time.

Why it’s interesting

Section titled “Why it’s interesting”

Nucleus-Image is the first MoE-image-generation announcement filed at Luma that frames itself explicitly around sparsity as the efficiency lever in T2I, rather than distillation, quantization, or architectural simplification — distinct from the dense T2I efficiency stories in PRX Part 3 — Training a Text-to-Image Model in 24h (PRX) and Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer (Z-Image). It is the image-modality counterpart to HunyuanImage 3.0 Technical Report‘s 80B-total / 13B-activated MoE — same architectural recipe, much smaller scale (17B vs 80B total; 2B vs 13B active), and the claim is that the smaller MoE can match dense frontier models from pre-training alone. If the release ships as promised (Apache 2.0 + training code + dataset recipe), it would be a directly auditable counterexample to the “MoE is for LLMs, dense is for diffusion” pattern, and a useful comparison point for Luma’s own scaling tradeoffs. Caveats worth flagging: no tech report at announcement, benchmarks are model-card-style aggregates that have been gamed before, and OneIG-Bench’s beating-Imagen-4 claim contradicts most independent T2I leaderboards — wait for weights + an independent run.

See also

Section titled “See also”