Krea Realtime 14B: Real-Time, Long-Form AI Video Generation

TL;DR

Krea Realtime 14B is a 14B-parameter autoregressive video model distilled from Wan 2.1 14B T2V via Self-Forcing — over 10× larger than prior open real-time video models, which were all based on Wan 2.1 1.3B. It achieves 11 FPS at 4 inference steps on a single B200, with ~1 s first-frame latency, and supports mid-stream prompt changes and live restyling. The release introduces two novel inference techniques for long-form stability — KV Cache Recomputation with block-causal masking, and KV Cache Attention Bias — plus FSDP/gradient- checkpointing/dynamic-KV-freeing tricks to make 14B Self-Forcing training fit in memory on 64–128 H100 GPUs.

Key claims

Krea Realtime 14B is distilled from Wan 2.1 14B (via LightX2V’s timestep-distilled checkpoint) using the three-stage Self-Forcing pipeline: bidirectional timestep distillation → causal ODE pretraining (block-causal mask) → Self-Forcing Distribution Matching Distillation [§“Training Krea Realtime”].
Achieves 11 FPS text-to-video at 4 inference steps on a single NVIDIA B200, with ~1 s first-frame latency [§“TL;DR”].
Causal ODE pretraining uses timestep shift 5.0 (vs. CausVid’s default 8.0) — lower shift preserves texture and dynamic motion at the lower SNR characteristic of lower-resolution video [§“Causal ODE Pretraining”].
Filters the original CausVid VidProm prompt dataset to remove low-diversity / static-motion prompts and resynthesizes prompts optimized for dynamic motion; prompt diversity is observed to affect downstream motion dynamism [§“Causal ODE Pretraining”].
Self-Forcing DMD training of a 14B student requires 4× 14B models in memory (real-score, critic-score, student, EMA), plus autoregressive student rollout plus a 25 GB-per-GPU KV cache — naive FSDP (ZeRO-3) sharding on 64 H100s still OOMs [§“Self-Forcing DMD”].
Dynamic KV Cache Freeing: after backprop through frame block N, its keys and values can never be needed again (block-causal structure), so backprop hooks on self-attention layers free per-transformer-block KV memory during the backward pass; this plus gradient checkpointing + FSDP makes 14B Self-Forcing tractable [§“Dynamic KV Cache Freeing”, Fig. 4].
KV Cache Recomputation (novel): on sliding-window eviction, instead of dropping the oldest block’s KVs, recompute KVs for the most recent N latent frames from their clean latents under a block-causal mask. This simultaneously fixes (a) the first-frame VAE distribution shift (the first RGB frame of a sequence encodes to a single-frame latent with different statistics from subsequent 4-frame-compressed latents) and (b) the receptive-field leakage that propagates errors from evicted frames into fresh KVs [§“KV Cache Recomputation: A Unified Solution”].
Recomputation cost is O(S² · N(N+1)/2) vs. vanilla Self-Forcing’s O(S²N), and requires Flex Attention (no Flash Attention 4 / Sage Attention support) — retained anyway because it is the only technique they found that delivers stable long-form generation [§“KV Cache Recomputation”].
KV Cache Attention Bias (novel): a negative attention bias applied only to past-frame tokens during sampling (current-frame self-attention unmodified and no bias during KV recomputation) prevents repetitive-motion collapse and loosens the consistency constraint enough to follow complex prompt transitions (e.g. “a woman” → “a bird”) [§“KV Cache Attention Bias”].
The cheaper baseline — First-Frame Anchoring (never evict frame 0, roll the rest of the cache) — is “remarkably effective” at preventing collapse on its own, but still produces artifacts and limits how far long generations can deviate from the initial frame [§“First Frame Anchoring”].
3 latent context frames (12 RGB frames, or 9 with first-frame re-encoding) is the sweet spot on Wan 2.1 14B — shorter context cuts attention latency (the dominant cost) and reduces error accumulation, but limits long-range temporal modeling [§“Shortening the Context Window”].
Prompt-embedding interpolation during inference enables both intra-prompt diversity (interpolating variations of the same prompt) and mid-stream prompt changes (e.g. switching subjects); attention bias is required to make high-divergence transitions actually follow the new prompt rather than preserving consistency with the old subject [§“Enhancing Diversity With Prompt Interpolation”].
Acknowledged limitation: DMD’s reverse-KL term is mode-seeking, suppressing meaningful low-probability modes; samples taken at ~25% of DMD training have more dynamic camera motion than the final model — a diversity–quality tension inherited from vanilla DMD [§“Limitations”].
Future-work targets named explicitly: ADM (Adversarial Distribution Matching, arXiv 2507.18569), DMD 2 (data + GAN), and AAPT (Autoregressive Adversarial Post-Training, arXiv 2506.09350) — all of which add real-data or GAN signal to the Self-Forcing recipe [§“Future Work”].

Method

The pipeline is three-stage Self-Forcing scaled to a 14B base. Stage 1 reuses LightX2V’s open-source timestep-distilled Wan 2.1 14B (bidirectional, 4-step). Stage 2 is causal ODE pretraining: synthesize ODE solution trajectories from the undistilled teacher under a block-causal attention mask (all frames within a block attend to each other; all frames attend to all past frames), then regress the timestep-distilled student to match. Stage 3 is Self-Forcing DMD — the student autoregressively samples videos during training (matching the inference task), and a Distribution Matching Distillation objective with two bidirectional score-estimator models (real and fake) pushes the student distribution toward high-probability regions.

The 14B-scale memory work is the central engineering contribution. Three techniques compose: FSDP (ZeRO-3) sharding across 64–128 H100 GPUs, gradient checkpointing, and Dynamic KV Cache Freeing — the latter exploits block- causality (block N’s KVs are needed only while backpropagating through block N itself) to free per-transformer-block KV memory during the backward pass via attention-layer backprop hooks. They train Stage 2 for 3k steps on 128 H100s.

The long-form inference contribution is KV Cache Recomputation with block-causal masking: rather than dropping the oldest block’s KVs at sliding- window eviction, recompute the KVs for the most recent N clean latent frames under the block-causal mask, optionally re-encoding the first RGB frame as a single-frame latent (matching the VAE’s first-frame statistics). This breaks the receptive-field leakage path that otherwise propagates information from already-evicted frames into “fresh” KVs and amplifies errors. KV Cache Attention Bias (a negative bias on past-frame attention scores during sampling, omitted during recomputation) further dampens drift and prevents motion-repetition collapse.

Results

Headline operating point: 11 FPS text-to-video at 4 inference steps on a single NVIDIA B200, ~1 s first-frame latency, on a 14B parameter base. Long-form generation is reported qualitatively — the post shows that with naive sliding-window eviction, both the 1.3B Self-Forcing model and Krea Realtime 14B collapse; first-frame anchoring stabilizes them but still drifts; KV Cache Recomputation + Attention Bias produces stable long-form output. No quantitative VBench / motion-amplitude / drift numbers are reported; the release is engineering-and-recipe-forward rather than benchmark-forward.

The post also reports the qualitative progression across the three training stages (timestep-distilled → causal-ODE-pretrained → final Krea Realtime 14B) via side-by-side rollouts, showing the model progressively adapting from parallel denoising to streaming generation.

Why it’s interesting

This is the largest open Self-Forcing model to date — 10× the parameter count of every prior real-time autoregressive video release — and the first to report the engineering recipe (FSDP + gradient checkpointing + dynamic KV freeing) that makes Self-Forcing tractable at 14B. It complements Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (the underlying technique, applied here at 14B) and Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation (AAPT, the data-driven adversarial alternative Krea explicitly names as future work), and contrasts with Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation and Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation, which argue the bidirectional-teacher → AR-student ODE init in Self-Forcing has a structural frame-level injectivity gap that should be fixed upstream of the rollout-drift problem Krea tackles. Krea’s KV Cache Recomputation and Attention Bias are inference-time drift fixes that sit alongside the training-time fixes in Context Forcing: Consistent Autoregressive Video Generation with Long Context (Slow/Fast/Sink cache partitioning) and PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference (learned mid-cache compression). It also occupies the same product niche as MirageLSD: The First Live-Stream Diffusion AI Video Model (MirageLSD) and Lucy — Decart's real-time image-to-video DiT (lucy.decart.ai) (Lucy), but with the actual recipe and memory-management tricks published.