Text-to-Image Models Fail to Count Objects Past Ten
Five open models follow number words only up to ten objects; SD3.5-M and FLUX.1-dev ignore counts beyond that.
Stories per day · 30 days
4 RSS
Relatedimage-generation5diffusion4multimodal4reinforcement-learning4autoregressive3diffusion-models3open-weights3bria2distillation2diversity2fibo2halcyon2text-to-video2visual-verifier23d-assets1adversarial-distillation1attention-routing1benchmarks1
50 stories, newest first · one card per story
Stories tagged #text-to-image
VA-TDM boosts few-step diffusion models
VA-TDM uses velocity adversarial distillation to reach state-of-the-art four-step text-to-image and text-to-video generation.
MiniMax releases open-weight MiniMax-H3-Image for generation and editing
The checkpoint outputs about 16-megapixel images in 6 steps and accepts up to 16 reference images.
StepX-OPD merges distillation stages for faster diffusion models
StepX-OPD lets one student learn all capabilities and few-step sampling using its own rollout states, scoring highest among four-step methods on SD3.5-Medium and FLUX.1-dev while using 37% fewer GPU-hours.
SpecSlot cuts global attention cost in pixel-space transformers
Only the lowest-frequency slot per spatial anchor attends globally; higher-frequency detail stays local.
Intrinsic Distribution Matching improves few-step image distillation
At two network evaluations, IDM scored 94.25% on Chinese and 90.20% on English LongText-Bench.
Few-step image models raise unsafe output; DICE cuts it
On FLUX Schnell, DICE lowers the unsafe-generation rate to 0.046, versus at least 0.161 for prior erasure methods tested.
Paper splits visual generation into autoregressive knowledge and diffusion rendering
A 3B text-to-image model trained in 6,600 H100 hours scored 86 on GenEval and 87 on DPG-Bench.
Qwen-Image-PE learns visual design for text-to-image prompt enhancement
The released 9B model beats evaluated prompt-enhancement baselines, while the 397B variant gains 7.81 points on Qwen-Image-Bench.
SpectralCanary uses low-frequency traces to audit T2I dataset use
With SD3.5 across five datasets it averages 92.4% per-generation TPR at 1% FPR and mean ΔFID of 0.05.
T-DIG lets users steer image diversity with natural-language prompts
It keeps the image generator frozen and adds user-directed diversity to open and closed models without weight access.
Comfy-Org republishes Bria Fibo 2.0 weights for ComfyUI
Dev and edit-turbo weights are merged for ComfyUI, with int8 builds, under CC BY-NC 4.0.
Residual-space adaptation cuts forgetting in continual diffusion personalization
Restricting updates to layers identified by residual dynamics keeps performance while cutting trainable parameters by up to 72%.
Halcyon trains text-to-image models without language pretraining
The authors report Halcyon-1.6B beating a larger JiT baseline on gFD with 275 times less pretraining compute.
Halcyon trains text-to-image models without language pretraining
Authors say Halcyon-1.6B beats a JiT baseline on gFD with 275 times less pretraining compute.
Bria AI Fibo-2.0-dev text-to-image model trends on Hugging Face
The dev checkpoint has 1,115 downloads and 157 likes and is listed under an other license.
MFlow learns when pixel diffusion should predict images or velocity
A post-hoc ConvRefiner, trained with the backbone frozen, restores high-frequency image statistics while preserving FID.
MOSS-Image open text-to-image system targets dense text and layouts
A 5B diffusion transformer plus an agentic prompt module reports 96.3% average text coverage; weights are promised for release.
LILIE language-informed loss improves pixel-space text-to-image training
JiT models trained with LILIE cut GPIC gFD from 66.84 to 53.01 versus a pixel-only baseline.
OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling
OmniVerifier-C1 grades decomposed visual rules continuously and beats binary verifiers on image and world-model tasks.
Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models
MSR improves generative-model diversity by sampling corpus-derived attributes and using the model only as a renderer.
OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling
OmniVerifier-C1 scores decomposed visual rules continuously and outperforms binary verifiers on image and world-modeling tasks.
Routing Policy Optimization for Online Diffusion Reinforcement Learning
RPO uses attention-route exploration to align diffusion generators about 4x more efficiently than Flow-GRPO.
OpenMOSS-Team/MOSS-Image-Base — new model trending #30 on Hugging Face
OpenMOSS released MOSS-Image-Base, an open text-to-image model now trending on Hugging Face.
7