Unified Multimodal Models for Understanding and Generation
Single models that both read images and generate them, so editing and reasoning share one backbone.
Tier Grades
Tier A: Frontier AI labs
feasible
Frontier labs ship native image generation inside multimodal models.
Tier B: Indonesian AI labs
not feasible
Pretraining unified models is not feasible for an Indonesian lab.
Tier C: Institution units such as ITB
stretch
A university unit can study and fine-tune 7B to 8B open unified models.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
stretch
Delivery teams use them through APIs; building on them is a stretch.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeTraining compute is not in the sources read.
- DataNot in the sources read; BAGEL reports free-form image manipulation and future frame prediction.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 12 authors on the BAGEL paper (arXiv 2505.14683).
- Time horizonYear.
- Shelf-life signalsFrontier chat models ship native image output, Open unified models match specialist generators
- Pipeline roleTier A trains unified models; Tier C studies them; Tier D uses them via APIs for editing workflows.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
Emu3.5image generationexemplar
HunyuanImage 3.0image generationexemplar
BLIP3-o (8B)image generationexemplar
BAGEL (7B MoT)image generationexemplar
Nano Banana 2.1image generationproduced by
Nano Banana 2 Liteimage generationproduced bySenseNova-U1-A3B-MoTimage generationproduced byNano Banana 2 (Gemini 3.1 Flash Image)image generationproduced byNano Banana Pro (Gemini 3 Pro Image)image generationproduced byBLIP3o-NEXT-GRPO-GenEval (3B)image generationproduced by
Lumina-DiMOO w/ Self-GRPOimage generationproduced byHunyuanImage 3.0 Instructimage generationproduced byGemini Nano Banana 2.1image generationproduced byNano Bananaimage generationproduced byHunyuanImage 3.5 Previewimage generationproduced byLinks
R&D TopicUnified Multimodal Models for Understanding and Generation
ModelsBAGEL (7B MoT)ByteDance, image generationBLIP3-o (8B)Salesforce, image generationEmu3.5BAAI, image generationHunyuanImage 3.0Tencent, image generation+11 more
BAGEL (7B MoT)ByteDance, image generationBLIP3-o (8B)Salesforce, image generationEmu3.5BAAI, image generationHunyuanImage 3.0Tencent, image generation+11 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 15
Nano BananaGoogle, image generationproduced byNano Banana 2.1Google DeepMind, Google, image generationproduced byHunyuanImage 3.0Tencent, image generationexemplarHunyuanImage 3.0 InstructTencent, image generationproduced byHunyuanImage 3.5 PreviewTencent, image generationproduced byNano Banana 2 (Gemini 3.1 Flash Image)Google, image generationproduced byNano Banana Pro (Gemini 3 Pro Image)Google, image generationproduced byNano Banana 2 LiteGoogle, image generationproduced byGemini Nano Banana 2.1Google, image generationproduced byBAGEL (7B MoT)ByteDance, image generationexemplarBLIP3-o (8B)Salesforce, image generationexemplarBLIP3o-NEXT-GRPO-GenEval (3B)Salesforce, image generationproduced byPapers
- Chameleon: Mixed-Modal Early-Fusion Foundation ModelsChameleon Team, arXiv preprint, 2024 founding
- Emerging Properties in Unified Multimodal PretrainingChaorui Deng et al., arXiv preprint, 2025 state of the art
- Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal ModelChunting Zhou et al., arXiv preprint, 2024 state of the art
- Show-o: One Single Transformer to Unify Multimodal Understanding and GenerationJinheng Xie et al., ICLR 2025 (per arXiv comment), 2024 state of the art
- Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and GenerationChengyue Wu et al., arXiv preprint, 2024 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2505.14683 https://arxiv.org/abs/2505.14683
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Unified Multimodal Models for Understanding and Generation", https://aurion.semai.id/topics/unified-multimodal-understanding-and-generation/, records of October 2026.