Unified image generation has a long-standing engineering default: subject-driven generation, canny-to-image, and instruction editing each keep their own adapter. If one model is to cover all of them, a video backbone looks like a natural host — it already learns how frames change into each other. FrameForge (arXiv:2610.36545), posted on arXiv on Oct 27, is the newest take on this idea: freeze HunyuanVideo1.5, train a single LoRA, and cover all three tasks. Yijing Lin, the lead author, also headed RealGeneral. The authors submitted the paper to the Hugging Face community board only yesterday; it now sits at the top of the daily papers board with 89 upvotes.

The fixed frame gap is an overlooked assumption

Existing crossovers (RealGeneral, DRA-Ctrl, ChronoEdit) mostly pin the condition frame and the target frame to fixed temporal positions, so every sample shares one "frame gap". The paper's observation: this default does not hold. Editing usually changes a small part; subject-driven generation and canny-to-image can overhaul appearance, structure, or the whole scene. The paper's LPIPS distributions show the condition-target variation magnitudes differ clearly across tasks and stay wide within each task — a task label cannot say how much an individual sample should change. Their answer: do not hardcode the gap; let the model pick per input.

MoTE: three temporal experts, a router picks the gap

The implementation is a Mixture of Temporal Experts. Three temporal RoPE experts cover three virtual frame gaps with equivalent video-frame spans of 17, 69, and 121. A lightweight router reads the text, condition, and noisy-target representations and picks one expert per sample (top-1); a straight-through estimator keeps discrete routing trainable end to end, and a load-balancing loss prevents collapse. The routing analysis is the most interesting part: over 500 cases and 50 denoising steps, canny — a task with large visual changes — favors the long-interval expert early; subject-driven generation and editing start on the medium interval and shift to the short one as target content emerges, with editing switching earlier and harder. Routing does not tag tasks; it moves with the denoising process.

CoTR: one unsupervised "middle frame"

The other module handles "how to change". CoTR inserts a noise-initialized reasoning latent between the condition and the target, letting it attend to text, condition, and target to gather context step by step. During the intermediate training stage, a stochastic attention mask (probability 0.5) cuts the target's direct attention to text and condition, forcing information through this latent pathway. There is no intermediate-frame supervision anywhere — only the target-frame diffusion loss drives the learning. The ablation also shows more latents are not better: two or three degrade all three tasks, and one is enough.

The numbers: where it wins and where it loses (all paper-reported)

The training bill is moderate: backbone fully frozen, a rank-32 LoRA, 182M trainable parameters, 170K samples per task (510K total), 2 A800 GPUs for about 6 days. DreamBench subject-driven overall 0.739, the highest among the compared methods, 1.37% over ChronoEdit (0.729), while subject consistency 0.739 still trails ChronoEdit (0.757). On COCO2017 canny-to-image, SSIM 0.447 is the best in the table, 7.2% over EasyControl; FID 23.47 loses to ControlNet's 20.42 — the authors themselves note FID and SSIM measure different things and should be read together. Editing is the highlight: in the paper's table on GEdit-Bench, the open-source best overall is 6.12, and motion-change editing reaches 6.84, above closed-source Gemini 2.5 (6.72) and Seedream 4.0 (5.53), still below GPT-4o (7.81); overall it also trails the three closed-source systems (7.81-8.01).

The ablation crack: one fix is not enough, two together overtake

The most informative row of the ablation table: a shared-LoRA unified baseline drags editing from the task-specific 5.99 down to 4.94 — unification genuinely interferes. Adding MoTE alone recovers to 5.77, CoTR alone to 5.56; neither reaches 5.99. Together they hit 6.12 and overtake the task-specific score. "Picking the gap" and "how to change" are complementary jobs; fixing only one is not enough. Switching to a HunyuanVideo1.0 backbone keeps the direction (editing 4.38 to 5.00), so the idea does not depend on one backbone. Code and LoRA weights are released, under the Tencent Hunyuan Community License.

What is worth remembering is not "video models won again", but that "how much the condition should change" went from a hand-set prior to an input-dependent, routed quantity. In the paper's comparison table, every open-source method except ChronoEdit — also video-backed — scores between 1.13 and 3.12 on motion editing; what image-native stacks lack may be exactly this sense of change magnitude. When the 6.12-to-8.01 gap closes is a question for the next adapter.