v0.3.12

What's Changed

• Nvidia Cosmos 7B and 14B: text to video and image to video diffusion model support.
• New sampler: res_multistep
• ckpt/pt/etc.. files are now always loaded safely on pytorch 2.4 and above.
• Fix some cases of ancestral samplers not being deterministic.
• Support ascend npu by @ji-huazhong in #5436
• Add option to log non-error output to stdout by @webfiltered in #6243
• serve workflow templates from custom_nodes by @bezo97 in #6193
• Remove duplicate calls to INPUT_TYPES by @catboxanon in #6249
• Fix Hook Keyframe 'guarantee_steps' behavior and add 'sigmas' by @Kosinkadink in #6273
• Add kl_optimal scheduler by @blepping in #6206
• (fix): "verbose" argument by @bigcat88 in #6289
• Fix custom node type-hinting examples by @webfiltered in #6281
• Add missing model_options param to finalize_default_conds call by @Kosinkadink in #6296
• Fix unknown scheduler error handling in calculate_sigmas function by @blepping in #6280
• Fix temporal tiling for Tiled VAE decoder, remove redundant tiles. by @kvochko in #6306
• Update web content to release v1.6.14 by @huchenlei in #6312
• add fov and mask for load 3d node by @jtydhr88 in #6308
• Update web content to release v1.6.15 by @huchenlei in #6324
• Update web content to release v1.6.16 by @huchenlei in #6335
• Update web content to release v1.6.17 by @huchenlei in #6337
• Add update-frontend github action by @huchenlei in #6336
• Update CODEOWNERS by @yoland68 in #6338
• In inner_sample, change "sigmas" to "sample_sigmas" by @Kosinkadink in #6360
• Frontend Update: v1.6.18 by @huchenlei in #6368
• Document get_attr and get_model_object by @huchenlei in #6357
• fixed: robust loading comfy.settings.json by @ltdrdata in #6383
• Add pyproject.toml by @huchenlei in #6386
• Hooks Part 2 - TransformerOptionsHook and AdditionalModelsHook by @Kosinkadink in #6377
• Merge ruff.toml into pyproject.toml by @huchenlei in #6431
• (fix): load_extra_path_config: relative path not converted to a full path by @bigcat88 in #6395
• Rewrite res_multistep sampler and implement res_multistep_cfg_pp sampler by @pamparamm in #6462
• Add SetFirstSigma node by @catboxanon in #6459

New Contributors

• @ji-huazhong made their first contribution in #5436
• @bezo97 made their first contribution in #6193
• @kvochko made their first contribution in #6306

Full Changelog: v0.3.10...v0.3.12

v0.3.11

What's Changed

• Nvidia Cosmos 7B and 14B: text to video and image to video diffusion model support.
• New sampler: res_multistep
• ckpt/pt/etc.. files are now always loaded safely on pytorch 2.4 and above.
• Fix some cases of ancestral samplers not being deterministic.
• Support ascend npu by @ji-huazhong in #5436
• Add option to log non-error output to stdout by @webfiltered in #6243
• serve workflow templates from custom_nodes by @bezo97 in #6193
• Remove duplicate calls to INPUT_TYPES by @catboxanon in #6249
• Fix Hook Keyframe 'guarantee_steps' behavior and add 'sigmas' by @Kosinkadink in #6273
• Add kl_optimal scheduler by @blepping in #6206
• (fix): "verbose" argument by @bigcat88 in #6289
• Fix custom node type-hinting examples by @webfiltered in #6281
• Add missing model_options param to finalize_default_conds call by @Kosinkadink in #6296
• Fix unknown scheduler error handling in calculate_sigmas function by @blepping in #6280
• Fix temporal tiling for Tiled VAE decoder, remove redundant tiles. by @kvochko in #6306
• Update web content to release v1.6.14 by @huchenlei in #6312
• add fov and mask for load 3d node by @jtydhr88 in #6308
• Update web content to release v1.6.15 by @huchenlei in #6324
• Update web content to release v1.6.16 by @huchenlei in #6335
• Update web content to release v1.6.17 by @huchenlei in #6337
• Add update-frontend github action by @huchenlei in #6336
• Update CODEOWNERS by @yoland68 in #6338
• In inner_sample, change "sigmas" to "sample_sigmas" by @Kosinkadink in #6360
• Frontend Update: v1.6.18 by @huchenlei in #6368
• Document get_attr and get_model_object by @huchenlei in #6357
• fixed: robust loading comfy.settings.json by @ltdrdata in #6383
• Add pyproject.toml by @huchenlei in #6386
• Hooks Part 2 - TransformerOptionsHook and AdditionalModelsHook by @Kosinkadink in #6377
• Merge ruff.toml into pyproject.toml by @huchenlei in #6431
• (fix): load_extra_path_config: relative path not converted to a full path by @bigcat88 in #6395
• Rewrite res_multistep sampler and implement res_multistep_cfg_pp sampler by @pamparamm in #6462
• Add SetFirstSigma node by @catboxanon in #6459

New Contributors

• @ji-huazhong made their first contribution in #5436
• @bezo97 made their first contribution in #6193
• @kvochko made their first contribution in #6306

Full Changelog: v0.3.10...v0.3.11

v0.3.10

What's Changed

• Add temporal tiling support to video VAEs exposed through the VAE Decode (tiled) or VAE Encode (tiled) nodes, this makes 8GB vram GPUs viable for hunyuan video.
• Add PixArt model support by @city96 in #6055
• Replace print with logging by @huchenlei in #6138
• Adding upstream pytorch installation instruction for Intel GPUs to README.md by @qiacheng in #6069
• remove minimum step count for AYS by @TechnoByteJS in #6137
• fixed a bug in ldm/pixart/blocks.py by @zhangp365 in #6158
• Add oneAPI device selector for xpu and some other changes. by @simonlui in #6112
• [Bug] Use raw dir name to serve static web content by @huchenlei in #6107
• Enable External Event Loop Integration for ComfyUI [refactor] by @bigcat88 in #6114
• VAE calculations are now done in bf16 by default on Mac for lower memory usage.

New Contributors

• @qiacheng made their first contribution in #6069
• @TechnoByteJS made their first contribution in #6137
• @zhangp365 made their first contribution in #6158

Full Changelog: v0.3.9...v0.3.10

v0.3.9

What's Changed

• LTXV 0.9.1 support (new VAE architecture)
• Tiled VAE encode node now works with video VAEs
• Fix a memory issue that lead to OOM when using a partially loaded model multiple times in a row.
• Support preview images embedded in safetensors metadata by @catboxanon in #6119

Full Changelog: v0.3.8...v0.3.9

v0.3.8

What's Changed

• Hunyuan Video Support: Example workflows
• Improve Timestep ranges on rectified flow models. (Potentially breaking change for workflows that use nodes with start/end percent on rectified flow models)
• Experimental support for sage attention with: --use-sage-attention, make sure to install it first.
• Add support for attention masking in Flux by @Slickytail in #5942 (better way of reducing the strength of flux redux, use strength_type attn_bias in "Apply Style model" node.
• Update web content to release v1.5.18 by @huchenlei in #6075
• Update web content to release v1.5.19 by @huchenlei in #6105
• ltxv: add noise to guidance image to ensure generated motion. by @michaellightricks in #5937
• Add MaHiRo (improved/alternate CFG) by @yoinked-h in #5975
• Add code owners by @huchenlei in #5873
• Fix SDXL generating black images when using --fast (fp8_ops) by @Haoming02 in #5928
• Lint unused import by @huchenlei in #5973
• Add cond_scale to sampler_post_cfg_function by @catboxanon in #5985
• Replace pylint with ruff by @huchenlei in #5987
• Update CI workflow to remove Windows testing configuration by @yoland68 in #6007
• Optimize model library by @hayden-fr in #5841
• Update README (Add new keybinding entries) by @huchenlei in #6020
• Lint all unused variables by @huchenlei in #5989
• Lint and fix undefined names (2/N) by @huchenlei in #6029
• Lint and fix undefined names (3/N) by @huchenlei in #6030
• Lint and fix undefined names (1/N) by @huchenlei in #6028
• Enforce F821 undefined-name by @huchenlei in #6032
• Enforce all pyflake lint rules by @huchenlei in #6033
• [Security] Fix potential XSS on /view by @huchenlei in #6034
• add load 3d node support by @jtydhr88 in #5564
• add_model_folder_path: ensure unique paths by removing duplicates by @bigcat88 in #5998
• fix: prestartup script is not applied due to extra_model_paths.yaml and ensure custom paths are used during startup by @ltdrdata in #5872
• (fix): added "model_type" to photomaker node by @bigcat88 in #6047
• use String directly to set bg color for load 3d canvas by @jtydhr88 in #6057
• Update README.md by @zlobniyshurik in #6071
• add preview 3d node by @jtydhr88 in #6070
• Add .github folder to maintainer owner list by @huchenlei in #6027
• Make CLIP set last layer node work with T5 models
• Add ConditioningStableAudio node to set some stable audio specific model conditioning.

New Contributors

• @michaellightricks made their first contribution in #5937
• @Haoming02 made their first contribution in #5928
• @catboxanon made their first contribution in #5985
• @hayden-fr made their first contribution in #5841
• @jtydhr88 made their first contribution in #5564
• @zlobniyshurik made their first contribution in #6071

Full Changelog: v0.3.7...v0.3.8

v0.3.7

What's Changed

• ModelPatcher Overhaul and Hook Support by @Kosinkadink in #5583
• Improved memory management. by @comfyanonymous in #5450
• Add rescaling_scale from STG to SkipLayerGuidanceDiT
• resolve relative paths in YAML configuration for extra model paths by @bigcat88 in #5847
• Add Create Hook Keyframes Interp. Node by @Kosinkadink in #5896
• [Developer Experience] Add node typing by @huchenlei in #5676
• Reland union type by @webfiltered in #5900

New Contributors

• @bigcat88 made their first contribution in #5847

Full Changelog: v0.3.6...v0.3.7

v0.3.6

What's Changed

• fast previews for LTXV
• Cropping can now be disabled in CLIPVisionEncode node to work around flux redux not working well on non square images.
• Add a basic multiply strength mode to the style model apply node as a rudimentary way to control the strength of flux redux.
• dpm_2_ancestral sampler now works on flow models (Flux, SD3.5, mochi, ltxv, etc...)
• fix multi add makedirs error by @lyksdu in #5786
• Remove internal model download endpoint by @huchenlei in #5432
• Update web content to release v1.4.13 by @huchenlei in #5807
• fix: backward compatibility patch for InpaintModelConditioning.encode by @ltdrdata in #5825

New Contributors

• @lyksdu made their first contribution in #5786

Full Changelog: v0.3.5...v0.3.6

v0.3.5

What's Changed

• SD3.5 Controlnet support.
• Fix issue with Flux fill (inpaint) model not working on some GPUs.
• Support official mochi loras.
• Skip layer guidance now works on LTX-Video and all other DiT models supported by ComfyUI.
• Some memory fixes with the lowvram system.
• Change EmptyLTXVLatentVideo minimum length from 9 to 1 by @spacepxl in #5750
• add better styling to README.md by @40476 in #5707

New Contributors

• @spacepxl made their first contribution in #5750
• @40476 made their first contribution in #5707

Full Changelog: v0.3.4...v0.3.5

v0.3.4

• Fix Tiled VAE decoding on LTX-Video
• Fix some lowvram mode issues with LTX-Video

Full Changelog: v0.3.3...v0.3.4

v0.3.3

• Fix issues with the lowvram system.

Full Changelog: v0.3.2...v0.3.3