v0.14.0

What's Changed

🎬 StreamDiffusion Pipeline — Complete Overhaul

• Prompt & Seed Blending: Smooth interpolation between values
• ControlNet Support: Depth, canny, and other controlnets with dynamic scale adjustment
• IPAdapter: Style transfer from reference images. FaceID version for face-preserving generation
• SD1.5 + SDXL: Expanded model architecture support
• NSFW Detection: Freepik-based safety checker
• Dynamic Parameters: In-place updates for most parameters without requiring pipeline restart (eg t_index_list, controlnets, etc)
• Dynamic Resolution TensorRT: Single engine with support for different resolutions (still requires a reload)
• TemporalNet ControlNet: Optical flow-based conditioning for frame-to-frame temporal consistency
• Multi-stage pipeline: Configurable image or latent, pre and post-processing in the pipeline (e.g.: latent feedback, upscaling, skip diffusion altogether)
• StreamV2V: Cached attention for even further temporal consistency

🔧 Realtime Runtime

• Extensible runner — Modular pipeline architecture with plugin pattern (#866)
• Loading screen — Visual feedback during pipeline reloads (#739)
• Removed Conda and moved to uv (except ComfyUI) (#854)
• Higher performance and lower latency
• Improve error handling and monitoring

📦 Batch Pipelines

• Audio-to-Text: Added faster-whisper backend (#595)
• LLM: Bumped vLLM version (#705)

New Contributors

• @vcashwin made their first contribution in #819
• @corey-livepeer made their first contribution in #837

Full Changelog: v0.13.8...v0.14.0

0.13.8

New features & improvements

• Support for audio passthrough and Opus encoding (real-time)
• Faster and more reliable startup & teardown (real-time)
• Improved logging, metrics, and debugging (real-time)
• Zero-copy GPU tensor transfer for lower latency (real-time)
• Streamdiffusion pipeline improvements and re-introduction (real-time)

Performance & stability

• Many encoder and decoder fixes (real-time)
• Better pipeline cleanup and failover handling (real-time)
• Robust support for multi-stream and resolution changes (real-time)
• Docker image optimizations and workflow improvements

Developer & maintenance updates

• Moved ai-worker package to go-livepeer
• Updated LLM pipeline dependencies to improve performance and fix bugs (batch)
• Improved CI workflows and build caching
• Enhanced checkpoint and model management

New contributors

• @Radovenchyk — first contribution [#417](#417)
• @livepeer-robot — first contribution [#611](#611)

Full Changelog: v0.13.3...v0.13.8

v0.13.3

What's Changed

• Support for JSON string in defaultImage field for overriding pipeline specific images in the mappings by @RUFFY-369 in #293

Full Changelog: v0.13.2...v0.13.3

v0.13.2

What's Changed

• examples: Add test run for lv2v by @victorges in #283
• dl_checkpoints.sh - docker run as a user not the root by @pwilczynskiclearcode in #397
• dl_checkpoints.sh run as root but chown to current user by @pwilczynskiclearcode in #399
• refactor(runner): update LLM 'tokens_used' to usage by @rickstaa in #402

Full Changelog: v0.13.1...v0.13.2

v0.13.1

What's Changed

• feat: update A2T audio conversion by @ad-astra-video in #389
• Remove liveportrait by @leszko in #393
• feat: better LLM response format by @kyriediculous in #387
• Fix codegen and some small updates by @ad-astra-video in #395

Full Changelog: v0.13.0...v0.13.1

v0.13.0

This release introduces several new ComfyUI pipelines, including Florence 2, LivePortrait, and Sam2, alongside significant performance optimizations and bug fixes for both the realtime AI stack and non-realtime pipelines like A2T and LLM. These updates enhance processing efficiency, stability, and functionality across the platform.

What's Changed

• Add nvidia-ml-py to requirements by @eliteprox in #305
• runner/live: Refactor pipeline streamer and add some docs by @victorges in #306
• Fix missing nvidia-ml-py by @leszko in #310
• runner/lv2v: Always exit main process when infer.py finishes by @victorges in #312
• (chore) fix startup with asyncio.create_task by @eliteprox in #314
• change github build action to run on oxford for live-pipeline builds by @pschroedl in #318
• runner/live: Ignore keepalive messages from control subscriber by @j0sh in #313
• Live video to video/comfyui sam2 by @pschroedl in #316
• ci(runner): add runner test ci by @rickstaa in #295
• ci(runner): add runner test ci by @rickstaa in #296
• ci(runner): change live pipelines docker runner by @rickstaa in #320
• feat: add flash_attention_2 to audio_to_text pipeline's attn_implementation by @RUFFY-369 in #308
• ci(worker): update go version by @rickstaa in #321
• ci(worker): fix worker test ci go version error by @rickstaa in #322
• runner/req: Add pynvml dependency by @victorges in #323
• Revert "(chore) fix startup with asyncio.create_task" by @eliteprox in #317
• fix: slow resize by @varshith15 in #315
• runner/live/trickle: Fix publisher/subscribe lifecycle by @victorges in #324
• (chore) remove cleanup-runner for self-hosted runners by @eliteprox in #328
• ci(runner): temporarily disable pipeline concurrency cancellation by @rickstaa in #329
• worker: Fix typo breaking build by @victorges in #332
• Init prompt from params by @pschroedl in #327
• ci(runner): remove cancel in context comment by @rickstaa in #333
• ci(runner): prevent pipeline docker CI from running on mac by @rickstaa in #331
• ci(runner): add amd64 runner label to pipeline docker ci by @rickstaa in #336
• ci(runner): disable fail fast for batch docker ci by @rickstaa in #337
• runner/docker: Fix streamdiffusion base image by @victorges in #340
• runner/docker: Install pynvml on noop by @victorges in #341
• Fix dl_checkpoints.sh: don use -it flag when running docker containers by @pwilczynskiclearcode in #339
• bump commit hash for streamdiffusion node, update dl_checkpoints.sh by @pschroedl in #344
• feat: Add florence 2 to comfyui container by @ad-astra-video in #343
• runner/live: Report pipeline status on trickle channel by @victorges in #309
• Update dl_checkpoints.sh for ComfyUI pipeline by @leszko in #346
• runner/docker: add cv2 to noop by @victorges in #348
• runner: Fix Awaitable import by @victorges in #350
• streamer: Fix remaining typos by @victorges in #351
• Update Sam2 ComfyUI Node by @pschroedl in #335
• feat: update florence 2 to add run mode by @ad-astra-video in #347
• ci(runner): enable docker build cancel-in-progress on feature branches by @rickstaa in #338
• dl_checkpoints.sh --tensorrt - exit if previous build is still running by @pwilczynskiclearcode in #357
• dl_checkpoints.sh new --batch flag to install only sam2-hiera-large model by @pwilczynskiclearcode in #358
• hotfix/resize_bug by @varshith15 in #360
• runner/pipelines: Accept ComfyUI prompt as a string by @victorges in #353
• streamer: Improve timestamp logic on status reporting by @victorges in #354
• runner/streamer: Fix nested process restart by @victorges in #361
• Fix model cache directory by @leszko in #365
• chore: update CODEOWNERS by @rickstaa in #366
• runner/docker: integrate ComfyUI LivePortrait by @emranemran in #342
• runner/live: Improve pipeline status schema by @victorges in #362
• update: Florence 2 commit for optimizations by @ad-astra-video in #363
• update: florence 2 node commit hash for bug fixes by @ad-astra-video in #369
• update comfui_stablediffusion, add checkpoints and lora by @pschroedl in #364
• streamer: Fix timestamp serialization as int ms by @victorges in #370
• Pull runner images by @leszko in #372
• dl_checkpoints.sh run correct docker image based on envs or default by @pwilczynskiclearcode in #374
• dl_checkpoints.sh - restore docker pull for mutable docker tags by @pwilczynskiclearcode in #375
• trickle: Add reconnects in subscriber. by @j0sh in #376
• Add a notification step after docker build succeeds for ai-runner:latest by @hjpotter92 in #377
• runner: Improve live video-to-video runner stability by @victorges in #371
• update florence2 models download by @ad-astra-video in #379
• live/streamer: Fix streamer shutdown logic by @victorges in #380
• docs: Add general documentation on realtime architecture by @victorges in #381
• Use vLLM to load LLMs by @kyriediculous in #230
• runner: fix openapi gen, add llm to workflows and container list by @ad-astra-video in #384

New Contributors

• @RUFFY-369 made their first contribution in #308
• @varshith15 made their first contribution in #315
• @pwilczynskiclearcode made their first contribution in #339

Full Changelog: v0.12.6...v0.13.0

v0.12.6

What's Changed

• live/runner: Increase liveportrait restart threshold by @victorges in #294
• Allow running worker on non-GPU machine by @leszko in #291
• fix: prevents nvml runtime error by @rickstaa in #292
• api: Fix startup of control API by @victorges in #300
• Add error handling to control message handler task by @eliteprox in #302
• runner: add events url to live video to video by @gioelecerati in #303
• runner: Make control_url and events_url optional by @victorges in #304
• SAM2 video-to-video real-time pipeline by @eliteprox in #280
• fix: managed containers return after inference completed by @ad-astra-video in #301

New Contributors

• @gioelecerati made their first contribution in #303

Full Changelog: v0.12.5...v0.12.6

v0.12.5

What's Changed

• feat: add hardware info reporting by @ad-astra-video in #273
• feat(worker): auto pull images if not found by @rickstaa in #200

Full Changelog: v0.12.4...v0.12.5

v0.12.4

What's Changed

• docker: ComfyUI model downloading/caching by @victorges in #284
• worker,runner: Fix lifecycle management of realtime containers by @victorges in #288
• test: add b64 functions tests by @rickstaa in #285
• worker: Add noop image to map by @victorges in #289
• Fix control_url task runner by @eliteprox in #290

Full Changelog: v0.12.3...v0.12.4

v0.12.3

What's Changed

• runner: Create liveportrait pipeline container by @victorges in #278
• worker+runner: Dynamically start live inference runners by @victorges in #275
• .github: Create CI for live pipelines docker images by @victorges in #282

Full Changelog: v0.12.2...v0.12.3