SolarWM is a fully open training stack for interactive video Video World Models that adapts four different pretrained video generators (5B–33B params) with one shared three-stage recipe, producing causal models that roll out coherent video for hours despite being trained only on 5-second clips.
Suppose you’re shipping a product that turns a still image into an explorable video scene the user steers with a virtual camera. Today you’d pick one video generator, hand-build a data pipeline against its quirks, and hope 5-second training clips somehow extend to minute-long interaction. Every time a stronger open video model drops (Wan, LTX, MiniMax), you start over. The dominant approach in the open world-model space, exemplified by systems like SANA-WM, releases one model tied to one backbone with partial data recipes. SolarWM’s pitch is that the pipeline around the base model, data normalization, camera conditioning, causal adaptation, is what actually determines whether interactive rollout works, and that this pipeline can be shared across backbones instead of rebuilt each time.
There are two pieces: a data engine and a training recipe, and they’re deliberately decoupled.
The data engine takes 10 heterogeneous source datasets (real-world video, synthetic, game footage) and forces every clip into one schema: video frames, metric-scale camera pose per frame, camera intrinsics per frame, a dense caption, quality metrics, and provenance. A key move: expensive preprocessing (camera estimation via a VIPE SLAM-based SLAM pipeline, captioning with Kimi K2.6, quality scoring) is done once per clip, and filtering thresholds live in a separate recipe layer. Rejected clips are kept with their rejection reasons, so changing a filter doesn’t mean re-running captioning. Dynamic entities that confuse camera-controlled models get scrubbed via a separate LTX Clean Plate pass that regenerates the video with people and vehicles removed.
The training recipe has three stages applied identically to all four backbones:
# Stage 1: bidirectional adaptation
# Teach the pretrained video model to obey camera conditions,
# with full temporal attention (no causality yet).
loss = mse(model(z_noisy, t, text, image, camera), flow_target)
# Stage 2: teacher-forced AnyFlow autoregressive init
# Split latents into ordered blocks. Predict block k given
# CLEAN ground-truth blocks < k. Switch to causal attention.
# AnyFlow loss trains few-step sampling directly.
loss = anyflow_loss(model, z_t_block_k, clean_history)
# Stage 3: DMD-based causal training
# Student rolls out on its OWN generations (matching inference).
# Frozen bidirectional teacher scores the target distribution;
# a trainable "fake" model scores the student's distribution.
# Their score difference pushes the student toward the teacher.
rollout = student.generate(init_image, camera_traj) # detached
grad = w(t) * (s_fake(rollout) - s_real(rollout))
Stage 2 is the trick that saves compute: prior recipes like Causal Forcing need separate ODE and consistency-distillation warmups before Distribution Matching Distillation (DMD). Teacher-Forced AnyFlow collapses those into one short pass because the bidirectional model from Stage 1 already knows appearance and motion; Stage 2 only has to teach it causal prediction. Camera control uses Fused Projective RoPE: instead of a separate camera branch, projective rotations from the camera pose are applied directly to the query, key, and value tensors inside the existing self-attention, so no extra attention pass is added.
The prevailing assumption in long-horizon video generation is that hour-scale rollout needs long training sequences, specialized attention-sink tricks, or fresh consistency-distillation runs per backbone. This paper argues the opposite. If data normalization, camera conditioning, and staged distillation are aligned properly, 5-second training generalizes to hour-scale inference across four unrelated video backbones without long-video fine-tuning or attention sinks. The evidence that carries this claim isn’t a benchmark score. It’s the demonstration that the same three-stage recipe produces working hour-scale causal generators on Wan 2.2, LTX-2.5, and MiniMax-H3 backbones from a single shared data pipeline.
The load-bearing result is a training-efficiency claim, not a leaderboard number. Three findings anchor it:
•
No specialized initialization is needed. State-of-the-art quality is reached without the separate ODE or consistency-distillation warmup stages that Causal Forcing-style pipelines require. Stage 2’s Teacher-Forced AnyFlow absorbs them.
•
Most compute belongs in bidirectional Stage 1. The paper reports that AR adaptation converges rapidly and Distribution Matching Distillation (DMD) needs even fewer steps, once Stage 1 is done well. This inverts the usual intuition that causal adaptation is the hard part.
•
Training on 5-second clips produces coherent hour-scale rollouts on all four backbones, with no attention-sink mechanism and no long-video fine-tuning. The SolarWM-wan2.2-5B-fast causal student sustains 60-minute uninterrupted generation from a single initial frame, with the scene prompt fixed and only the camera trajectory time-varying.
The released corpus itself is a secondary artifact: 1.43M canonical clips from 10 sources (~25.85 TB), of which 876k clips pass the default quality gates (471k “high,” 404k “xhigh”), and the remaining 549k rejected clips are still published with their rejection metadata so users can re-filter. Note: the paper does not report standard quantitative benchmark numbers (FVD, VBench, etc.) in the provided text. Evaluation is qualitative and horizon-based.
Reach for this when you’re building a camera-controllable video product and want to compare backbones fairly, or when you want to run your own data mixture without rebuilding annotation pipelines. Say you’re deciding between Wan 2.2 and LTX-2.5 for an image-to-explorable-video feature: SolarWM gives you both under the same data recipe and camera-conditioning interface, so a quality difference actually reflects the backbone rather than your integration effort. If you have your own proprietary video corpus, you can push it through the released engine, get frame-aligned camera geometry and captions, drop it into the mixture with a source weight, and retrain, without touching the expensive parts of preprocessing.
Artifacts: the dataset and code are released, along with model weights for all four family members (SolarWM-wan2.2-5B, SolarWM-wan2.2-14B, SolarWM-ltx-2.5-22B, SolarWM-minimax-h3-33B), the full data-processing pipeline, exact selection policies per source, training recipes for all three stages, and inference code. The 549k rejected clips ship with rejection reasons, so an alternative recipe can promote them back in. Licensing terms for the release aren’t stated in the paper text provided.
When a training pipeline needs to generalize across backbones, put the invariants in the data contract and the camera interface, and let each backbone keep its native representation. The reason SolarWM’s three-stage recipe transfers across four unrelated video generators isn’t that the recipe is clever, it’s that the interface around it (frame-aligned camera geometry, unified sample schema, Fused Projective RoPE injection into existing attention) is narrow enough not to fight any one backbone’s pretraining.
•
Evaluation is qualitative in the provided text. The paper shows minute- and hour-scale rollouts through sampled frames and describes them as “recognizable and visually coherent,” but doesn’t report FVD, VBench, or head-to-head numeric comparisons against prior interactive world models. “State-of-the-art” is claimed but not quantified in the excerpt.
•
Hour-scale coherence is a visual claim, not a semantic one. The model preserves scene layout and appearance under a fixed prompt and predetermined camera trajectory. Whether it maintains object permanence, spatial memory of visited regions, or physical consistency over an hour isn’t measured; the training signal is 5 seconds of context.
•
The “four backbones under one recipe” claim rests on the smallest backbone. The showcased long-horizon results (third-person, minute-scale, hour-scale, OOD initialization) are almost all generated by SolarWM-wan2.2-5B-fast. Whether the 14B, 22B, and 33B variants achieve equivalent hour-scale coherence is shown mainly for the earlier bidirectional stage, not the distilled causal one.