SenseNova U1.5-8B-MoT-Preview — Native 4K Generation and Editing on NEO-unify

ListenNo Encoder, No VAE, Native 4K10:57

Show notes

SenseTime's SenseNova U1.5-8B-MoT-Preview is an early open-source refresh of their encoder-free NEO-unify unified multimodal model — pixels and words in, pixels and words out, no vision encoder and no VAE — now doing native 4K text-to-image and reference-guided editing. We dig into the two changes that matter: a ConvDecoder patch-joint reconstruction head that kills grid seams at high resolution, and a cleaned-up editing corpus that improves identity preservation without a dedicated reference encoder. We read the benchmark deltas honestly, including where open encoder-free models still trail closed-source on multi-axis editing. **Sources** - SenseNova U1.5-8B-MoT-Preview — Native 4K Generation and Editing on NEO-unify

Subscribe to the channel: https://podcast.lumalabs.link/27a25f51ea81d41a8253f46c6133a35e/feed.xml

TL;DR

SenseNova U1.5-8B-MoT-Preview is an early open-source refresh of SenseTime’s natively unified multimodal model built on the encoder-free NEO-unify architecture (no vision encoder, no VAE — pixels and words are the direct I/O). At just 8B MoT parameters it targets native 4K text-to-image generation and reference-guided editing with quality that the team claims rivals commercial closed-source models. Two architectural deltas over U1: (i) a ConvDecoder patch-joint reconstruction head replacing U1’s independent-per-patch MLP head to suppress grid artifacts at high resolutions, and (ii) a heavily filtered/synthesized editing corpus with English–Chinese balance and multi-image reference coverage. Reported gains over U1: Qwen-Image-Bench 47.14 → 55.20 (with prompt enhance), ImgEdit-Bench 3.90 → 4.37, GEdit-Bench-en 7.47 → 8.17, GEdit-Bench-zh 7.42 → 8.05.

Key claims

The MLP head in the original SenseNova U1 reconstructs each RGB patch independently and at high resolutions exposes token boundaries as grid seams — U1.5 replaces it with a ConvDecoder that reshapes visual tokens into a 2D feature grid and upsamples through multiple Pixel Shuffle stages with intermediate 3×3 convolutions, enabling neighboring patches to interact and merge into a continuous image [§What U1.5 Improves — Patch-Joint Reconstruction].
The encoder-free NEO-unify architecture achieves strong instruction following together with subject-identity and structural preservation using only a single sequence of reference-image tokens; the team frames this as validating that the frozen understanding branch is sufficient to condition editing without a dedicated reference encoder [§What U1.5 Improves — Visual Preservation].
Despite training only on short JSON-formatted prompts (long/structured JSON appears only in the understanding corpus), U1.5-Preview generalizes to long and structured instructions at inference — treated as evidence that native unified training transfers structured comprehension from understanding to generation without dedicated prompt-template training [§What U1.5 Improves — Cross-Task Generalization].
Native 4K generation is claimed with “remarkable efficiency,” preserving both global coherence and fine-grained detail [§Overview].
Improvements over U1 on the shared benchmarks: Qwen-Image-Bench overall 48.28 → 49.93 (49.93→55.17 with prompt enhance, EN); ImgEdit-Bench 3.90 → 4.37; GEdit-Bench EN G_O 7.470 → 8.172; GEdit-Bench CN G_O 7.420 → 8.051; WeEdit average 6.497 → 6.852 [§Text-to-Image, §Image Editing].
U1.5-Preview with prompt enhance (Cosmos3-style upsampling) reaches Qwen-Image-Bench 55.17 EN / 55.22 ZH — competitive with GPT Image 1 (54.24 / 54.07) but below GPT Image 2 (65.23 / 64.69) [§Text-to-Image Qwen-Image-Bench Table].
Known limitations acknowledged: unintended text generation from short prompts, errors in dense/long text (small characters, mixed CN-EN), incomplete adherence to complex layouts (counts, alignment, hierarchy), instability in small faces/hands/limbs, and drift in complex multi-turn multi-reference edits [§Known Limitations].
Two operational recipes ship alongside the model — a Prompt-Enhance (PE) Skill that converts a short brief into a structured Render JSON prompt, and a Caption-then-Reconstruct pipeline for reference-image reinterpretation — both external to the raw model and not part of the base capability [§Best Practices].

Method

The core change is at the pixel-side head, not the backbone. NEO-unify’s U1 already replaced both the CLIP-style vision encoder and the Flux-style VAE with direct pixel I/O through what the team calls a “near-lossless visual interface” over a Mixture-of-Transformer backbone (understanding and generation pathways sharing the transformer infrastructure but routing through separate parameters — text supervised by AR cross-entropy, vision supervised by pixel flow matching). See NEO-unify: Building Native Multimodal Unified Models End to End for the underlying architecture.

U1.5-Preview leaves that backbone intact and rebuilds the decode head. The original U1 MLP head predicted each patch’s RGB independently; at high resolutions this leaves visible token-boundary seams. U1.5 replaces it with ConvDecoder: visual tokens are reshaped into a 2D feature grid, then progressively upsampled through multiple Pixel Shuffle stages interleaved with 3×3 convolutions that let neighboring patches exchange information before the final RGB projection. The team reports that this design “substantially suppresses grid-like artifacts throughout pre-training” and is more robust to downstream fine-tuning than the independent-patch head.

The editing side is a data change rather than an architectural one. U1 trained on largely public editing corpora with noisy supervision and visible compositing artifacts. U1.5 extensively cleans, filters, and synthesizes the editing data, improves EN–CN balance, and expands coverage of single-image and multi-image reference settings. Reference-image conditioning still flows through the same “single sequence of reference-image tokens” routed to the frozen understanding pathway — the team frames the U1.5 gains as evidence that the encoder-free channel is sufficient for identity and structural preservation without a dedicated reference encoder.

The reference inference config is cfg_scale=4.0, timestep_shift=3.0, num_steps=50. Pre-training launcher and configuration are open-sourced alongside the weights.

Results

Qwen-Image-Bench (Overall EN / ZH): U1 48.28 / 45.99 → U1.5-Preview 49.93 / 50.25 → U1.5-Preview with prompt enhance 55.17 / 55.22 [§Qwen-Image-Bench Table]. Reference points on the same table: GLM Image 47.86 / 48.19, Qwen Image 48.48 / 49.23, HunyuanImage 3.0 51.35 / 50.81, Qwen Image 2512 51.32 / 52.06, GPT Image 1 54.24 / 54.07, Qwen Image 2 (team’s re-eval) 55.69 / 55.63, GPT Image 2 65.23 / 64.69.
ImgEdit-Bench Overall: U1 3.90 → U1.5-Preview 4.37, sitting between Qwen-Image-Edit-2511 (4.51) and Nano-Banana-Pro (4.37), below FireRed-Image-Edit (4.56) [§ImgEdit / GEdit Table].
GEdit-Bench EN G_O / CN G_O: U1 7.470 / 7.420 → U1.5-Preview 8.172 / 8.051; best reported on the table are Qwen-Image-2 at 8.369 / 8.348 [§ImgEdit / GEdit Table].
WeEdit average (IA + TC + BP): U1 6.497 → U1.5-Preview 6.852; best reported on the same table are Nano-Banana-Pro 8.843, followed by GPT-Image-1.5 6.817 and Nano-Banana 6.287. U1.5-Preview leads all evaluated open-source competitors (FireRed-Image-Edit 5.873, HY-Image-3-Instruct 5.727, Qwen-Image-2 5.601, LongCat-Image-Edit 5.230, Qwen-Image-Edit-2511 3.913) but sits well below the closed-source Nano-Banana-Pro on this composite editing metric [§ImgEdit / GEdit Table].
The Preview is explicitly framed as an early open-source look; a “more capable and refined production version” (U1 Pro) is teased for imminent release. Prompt-enhance (PE) numbers rely on an external Skill invoking a rewriter model — not part of the raw model’s capability [§Best Practices].

Why it’s interesting

U1.5-Preview is the second architectural iteration of an encoder-free unified multimodal model to reach the wiki with matched-benchmark comparisons against U1, and both changes are diagnostic. The ConvDecoder swap directly targets a specific failure mode of the “no VAE” approach — pixel-space heads have to solve tile continuity themselves without an autoencoder to smooth over patch boundaries — and the reported artifact suppression is a concrete signal that the NEO-unify direction can scale to 4K without giving up the encoder-free property. This complements NEO-unify: Building Native Multimodal Unified Models End to End‘s original 2B / 90K-step preview by showing the same architectural family holding up at 8B and at production resolutions. The editing story is orthogonal but sharper for the wiki’s Unified Multimodal Models debate: U1.5’s identity/structure preservation improvements come purely from data curation on top of the same encoder-free reference-token channel, arguing that a dedicated reference encoder isn’t structurally required and contrasting with the “frozen tower + trained connector” recipes documented in GLM-5.2-Vision-NVFP4 — bolting MoonViT onto frozen GLM-5.2 with a 49.5M projector. The WeEdit gap to Nano-Banana-Pro (6.85 vs 8.84) is worth flagging — closed-source editing still dominates the composite benchmark despite competitive per-benchmark numbers on ImgEdit / GEdit, so encoder-free open UMMs at 8B remain a step behind the frontier on multi-axis editing quality.