autoregressive chunk-based long-video generation from text prompts
Generates minute-scale videos (up to 60+ seconds) from natural language text prompts using a 14B-parameter diffusion model with autoregressive, chunk-based frame generation. The model processes video in 33-frame chunks sequentially, with each chunk conditioned on previous chunks to maintain temporal coherence without explicit anti-drifting mechanisms like self-forcing or error-banks. Achieves 19.5 FPS on a single H100 GPU by leveraging unified history injection and multi-term memory patchification during training.
Unique: Achieves minute-scale video generation without conventional anti-drifting strategies (self-forcing, error-banks, keyframe sampling) by using unified history injection and multi-term memory patchification during training, enabling simpler inference pipelines and faster generation on single-GPU setups.
vs alternatives: Faster than Runway ML or Pika Labs for long-form generation (19.5 FPS on H100) because it avoids expensive anti-drifting mechanisms through training-time optimizations rather than inference-time corrections.
image-to-video conditional generation with visual grounding
Generates videos conditioned on a static input image, using the image as a visual anchor to guide the diffusion process. The model encodes the input image through the same VAE and transformer backbone used for text conditioning, allowing the image to provide spatial and semantic constraints that shape frame generation across all 33-frame chunks. Supports both Helios-Base (highest quality) and Helios-Distilled (fastest) variants with identical architectural conditioning.
Unique: Uses unified VAE and transformer conditioning pathway for both text and image inputs, enabling seamless switching between T2V and I2V tasks without separate conditioning modules or architectural branching.
vs alternatives: More flexible than Runway's image-to-video because it supports the same three model variants (Base/Mid/Distilled) for I2V as T2V, allowing quality-speed tradeoffs that competitors don't expose.
unified history injection for temporal coherence without explicit anti-drifting
Training mechanism that injects previous chunk history (encoded representations of prior 33-frame chunks) directly into the transformer attention layers, enabling the model to maintain temporal coherence across chunk boundaries without explicit anti-drifting strategies like self-forcing, error-banks, or keyframe sampling. The history is injected as additional context tokens in the attention mechanism, allowing the model to learn implicit drift prevention during training. This approach simplifies inference (no need for complex anti-drifting logic) while maintaining quality across minute-scale videos.
Unique: Injects previous chunk history as additional context tokens in transformer attention rather than using separate anti-drifting modules, enabling implicit drift prevention learned during training rather than explicit inference-time corrections.
vs alternatives: Simpler than self-forcing or error-bank approaches because it requires no inference-time logic — drift prevention is entirely baked into model weights, reducing inference complexity and latency.
easy anti-drifting training strategy for motion stability
Training-time technique that applies lightweight anti-drifting constraints during the Base model training stage, preventing motion drift without the computational overhead of inference-time anti-drifting mechanisms. The strategy uses multi-term memory patchification to reference multiple previous chunks, enabling the model to learn motion consistency across longer temporal windows. This is distinct from unified history injection — easy anti-drifting focuses on motion stability through explicit training objectives, while history injection provides implicit temporal context.
Unique: Applies anti-drifting constraints during training rather than inference, enabling lightweight motion stability improvements without the computational cost of inference-time mechanisms like self-forcing or error-banks.
vs alternatives: More efficient than inference-time anti-drifting because it bakes motion stability into model weights during training, avoiding the need for dual-pass inference or complex post-processing logic.
heliosscheduler and heliosdmdscheduler noise scheduling for variant-specific optimization
Two custom noise schedulers optimized for different prediction types and guidance strategies: HeliosScheduler for Base/Mid variants (v-prediction with standard/CFG-Zero guidance) and HeliosDMDScheduler for Distilled variant (x0-prediction with CFG-free guidance). Each scheduler is jointly optimized with its corresponding prediction type and guidance strategy during training, enabling faster convergence and better quality at fewer inference steps. The schedulers define the noise level progression across diffusion steps, with HeliosDMDScheduler using more aggressive noise reduction for x0-prediction.
Unique: Variant-specific schedulers (HeliosScheduler vs. HeliosDMDScheduler) are jointly optimized with prediction type and guidance strategy during training, enabling architectural adaptation rather than using a single universal scheduler.
vs alternatives: More efficient than fixed schedulers (e.g., linear, cosine) because each scheduler is co-trained with its prediction type and guidance strategy, enabling faster convergence and better quality at fewer steps.
video-to-video style transfer and motion continuation
Generates new video frames conditioned on an input video sequence, enabling style transfer, motion continuation, or video interpolation. The model encodes the input video through temporal convolutions and attention layers, extracting motion and semantic patterns that guide the diffusion process for subsequent frames. Supports frame-by-frame or chunk-by-chunk conditioning depending on the inference interface used.
Unique: Encodes input video through the same temporal transformer backbone used for training, extracting motion patterns without separate optical flow or motion estimation modules, enabling end-to-end differentiable video conditioning.
vs alternatives: Simpler than Deforum or Ebsynth because it doesn't require explicit optical flow computation or keyframe specification — motion is implicitly learned from the input video encoding.
progressive distillation pipeline with quality-speed tradeoff variants
Provides three model checkpoints (Helios-Base, Helios-Mid, Helios-Distilled) arranged in a distillation chain that progressively trades quality for inference speed. Base uses v-prediction with standard CFG and 50 inference steps for highest quality; Mid uses CFG-Zero with 20 steps per stage; Distilled uses x0-prediction with CFG-free guidance (scale=1.0) and 2-3 steps per stage. Each variant uses a different noise scheduler (HeliosScheduler for Base/Mid, HeliosDMDScheduler for Distilled) optimized for its prediction type and guidance strategy.
Unique: Distillation chain uses different prediction types (v-prediction → x0-prediction) and guidance strategies (Standard CFG → CFG-Zero → CFG-free) rather than just reducing model size or step count, enabling architectural adaptation at each stage rather than uniform compression.
vs alternatives: More transparent than Runway or Pika Labs because it exposes three distinct checkpoints with documented quality-speed tradeoffs, allowing developers to make informed variant selection rather than being locked into a single model.
multi-scale sampling pipeline with pyramid unified predictor
Helios-Mid and Helios-Distilled variants employ a multi-scale sampling pipeline that decomposes the diffusion process into multiple stages, each operating at different noise scales. The Pyramid Unified Predictor (PUP) architecture enables efficient coarse-to-fine generation where early stages produce low-frequency motion and semantic structure, and later stages refine high-frequency details. This approach reduces effective inference steps (20 per stage for Mid, 2-3 per stage for Distilled) while maintaining temporal coherence across chunk boundaries.
Unique: Pyramid Unified Predictor enables stage-specific prediction types and schedulers (v-prediction in early stages, x0-prediction in later stages) rather than uniform prediction across all diffusion steps, allowing architectural adaptation to noise scale.
vs alternatives: More efficient than standard multi-step diffusion because it uses a unified predictor across stages rather than separate models, reducing memory overhead while maintaining quality through hierarchical decomposition.
+5 more capabilities