Custom ComfyUI is now available on Comfy Cloud
Comfy officially announced that Comfy Cloud now supports custom ComfyUI setups (Custom Builds) for teams, allowing custom nodes, models, and LoRAs to run in the browser.
Comfy officially announced that Comfy Cloud now supports custom ComfyUI setups (Custom Builds) for teams, allowing custom nodes, models, and LoRAs to run in the browser.
ComfyUI workflow_templates v0.11.111 released, adding Kling 4.0 image, image editing, and long image templates, installable via pip install comfyui-workflow-templates==0.11.111. The templates were submitted by @comfyui-wiki, and the full changelog is v0.11.110...v0.11.111.
The author has released a fully automatic Viggle character motion transfer ComfyUI workflow that can run with 6GB of VRAM. The workflow relies on the INT8 ConvRot version of the Viggle model, and can also use a LoRA version that converts MiniMax H3 Ref2V into Viggle mode; the process only requires loading a character reference image, uploading a motion video, and selecting a duration, while the built-in Qwen VL prompt system analyzes the video and generates the result.
Seedance 2.5 Flash is now available in ComfyUI via Partner Nodes, supporting text-to-video, first/last frame, and multi-reference generation, with the ability to edit or extend reference videos, generating clips of 4-30 seconds, up to 720p, with native audio, covering aspect ratios from 9:16 to 21:9. The model outperforms Seedance 2.0 and costs less than Seedance 2.5.
The author tested RelayVSR 2x and 4x video super-resolution in ComfyUI. On an RTX 4070 Ti SUPER, a 5-second 0.5MP video took about 28 seconds at 2x and about 53 seconds at 4x.
The author shares a workflow for maintaining character consistency in ComfyUI: split the character description and scene description into two frozen blocks; the character block is never edited, while the scene block is changed for each shot. A fixed seed makes the armor texture and hair strands match. The recommended IPAdapter FaceID weight is 0.75-0.85; above 0.9 the character looks pasted on, and below 0.6 the face drifts.
The pDMD 2-step LoRA for MiniMax-H3-Image, using ComfyUI 0.45.0 and an RTX 4070 12GB, reduced the average generation time for 9 image edits from 38.83 seconds to 27.62 seconds, a reduction of about 28.9%.
A creator in Budapest used a single RTX 5090 to run MiniMax H3 locally in ComfyUI, completing a 2-minute 36-second sci-fi AI short drama pilot episode. Each shot was first previewed at 672x384 with four steps for screening, then formally rendered at 1344x768 with twenty steps, downscaled to 1080p via 2x super-resolution, with no cloud usage and no per-use charges throughout.
ComfyUI core adds a Video Deflicker node, merged via PR #17754, which can fix flicker after per-frame upsampling and other processing. The node requires downloading RAFT-large weights to ComfyUI/models/optical_flow/, using a stable frame as guide to read motion, preserving real changes. The same PR replaces torchvision raft_large with a native implementation, about 2.3x faster.
The open-source community has quantized MiniMax H3 Image to Q2_K, reducing VRAM requirements from 68GB to 6.4GB, allowing 8GB graphics cards to run 4K image generation locally. The quantized version natively outputs 4096×4096 and 21:9 widescreen, official distillation produces images in 6 steps, integrates generation and editing, and supports up to 16 reference images.
The author used MiniMax H3, LTX-2.5, and Wan 2.2 in ComfyUI to complete a 9:51 dialogue-free animated short film 'DAY 731', with most shots generated by H3.
LTX's VFX nodes are now available in ComfyUI via Partner Nodes, allowing post-processing to be completed without leaving the node graph. Supports recomposition (1:1, 4:5, 5:4, 9:16, 16:9), upscaling to 4K/8K, SDR to 10-bit HDR (BT.2020/HLG), and per-frame alpha generation up to 4K.
160 million views only brought in 390,000 yuan in revenue sharing, while cloud computing trial-and-error costs often range from tens of thousands to hundreds of thousands of yuan, making the pure cloud-based AI short drama business model untenable. The article presents a layered hybrid workflow—cloud-based scriptwriting, local image generation to lock character designs, low-resolution trial-and-error, and post-production super-resolution—which reduces the comprehensive cost per shot from over 20 yuan to a few jiao.
The author released the Geiru Window Director and Geiru Hybrid Windows workflow for single-pass rendering of MiniMax H3 long videos. Hybrid Windows renders long videos by overlapping windows, keeping VRAM usage to one window; Window Director supports per-window review, redo, and rollback, and can save each window to disk.
ComfyUI's new node H3-FaceRefine-Plus has been released, offering 2x detail on small and medium faces, developed based on Carasibana's Facerefine node. The author used Claude Code for development; the node is slower but provides richer detail, and the GitHub repository includes all workflows and resources needed for reproduction testing.
Seedream 5.5 is now available in ComfyUI via Partner Nodes, adding multi-round local editing, native transparent RGBA output, poster and image-text layout, more realistic portraits, and prompt-based lighting control. All features are accessible through a single node in any workflow. Users can try two entry points on Comfy Cloud: image editing and text-to-image.
MiniMax H3 Image receives day-0 native support in ComfyUI, with a single checkpoint supporting both text-to-image and reference image editing. The model is CFG-distilled, requires 6 sampling steps and no negative prompt, natively outputs resolutions such as 4096×4096 (1:1), and supports multi-reference image editing and prompts up to 4096 tokens.
ComfyUI tutorial uses RelayVSR and DLSS5 nodes to fix detail loss in MiniMax H3 video upscaling. Under 6GB VRAM, it can upscale a 0.7 MP video to 2K in about 7 minutes. Compared to MiniMax latent space upscaling which takes about 45 minutes, this method significantly reduces upscaling time.
The ComfyUI port of the ReCaVSR video super-resolution model has been released. int8 quantization allows 1440p 4x super-resolution to be completed on 8GB VRAM. The model is based on Wan2.2 TI2V-5B and comes with built-in FlashDecoder streaming decoding, so long clips only increase time and disk usage. In testing, a rainy street scene from 960×480 to 3840×1920 took about 36 seconds on an RTX 4090, with the int8 file at 6.27GB and output differing from bf16 by about 48dB.
ComfyUI's October update brings multiple efficiency improvements including branch nodes, LoRA stacking, automatic template downloads, multi-output zip export, EXR/HDR/ProRes support, link organization, and in-widget undo/redo, all included in v0.44. Branch Node supports up to 32 inputs, LoRA stacking node supports up to 20 LoRAs, and template downloads are limited to official templates.