Twelve 2026 papers show a 6–7B diffusion model can run on a pretrained LLM — Semantic Routing and LLaDA2.0-Uni fit the retrofit closest
Asked:“most promising research directions on integrating a ~7b txt2img diffusional model into a pretrained llm to do efficient img gen/editing? ideally throwing away og txt encoder”
Asked (follow-up):“just find and list recent 2026 papers on this that have promising results / direction”
Twelve 2026 arXiv papers (January 6 – October 9, 2026; revised older papers excluded), covering image models from 1.6B to 9.8B parameters conditioned by pretrained LLM/VLM backbones. Ranked by direct relevance to the proposed retrofit, not citation count; six have official GitHub repositories. All results are author-reported.
A · direct encoder replacement
B · unified LLM / diffusion architecture
C · connector or empirical baseline
D · efficiency / scaling follow-up
— vertical position is relevance rank (1 = closest match), horizontal is 2026 release date. Hover a dot for the full claim.
Semantic Routing is the purest version of the ask: it drops the old encoder for Qwen3-VL-4B and shows depth-wise mixing of LLM layers beats the standard penultimate-layer tap by +4.11 GenAI-Bench and +9.97 on Counting.
arxiv.org
UniDFlow posts the strongest numbers in the list — GenEval 0.95, DPG-Bench 91.19, ImgEdit 4.24 in 20 denoising steps — by repurposing a frozen Qwen3-VL as the discrete-flow generator itself.
arxiv.org
LLaDA2.0-Uni matches the 6–7B scale: it conditions a 6B pretrained diffusion decoder solely on semantic tokens from a diffusion LLM — no text prompt path at all — at GenEval 0.89.
arxiv.org ·
github.com
LLaDA-Image is the most reusable starting point: a 6B DiT with no conventional text encoder, a fully open training recipe, strong editing (GEdit-EN 7.336) and a few-step Turbo track.
arxiv.org ·
github.com
The full shortlist, ranked by relevance
| # | Paper | Class | Date | Scale / backbone | Editing | Headline result | Why it matters | Links |
|---|
Reading list of 12 papers released on arXiv between 2026-01-01 and 2026-10-18, assembled 2026; all 12 carry true 2026 arXiv identifiers (revised older papers excluded). Rank is editorial relevance to retrofitting a ~7B text-to-image diffusion model onto a pretrained LLM while discarding the original text encoder. Headline scores (GenEval, DPG-Bench, ImgEdit, GEdit, GenAI-Bench) are author-reported from each paper; benchmark protocols differ between papers, so scores are not directly comparable across rows. Official GitHub repositories exist for 6 of the 12; long abstracts and full result tables were cut for space.