Sign In
homemodelsimagesvideos3D Modelshubsarticlescomicschallengesupdatesshop

Recent ComfyUI Massive Improvements We Made for MiniMax H3 and More

0

SECourses's Avatar

SECourses

Forge Badge

Aug 27, 2026

announcement
Recent ComfyUI Massive Improvements We Made for MiniMax H3 and More

ComfyUI Improvements

Zip file that includes everything with full changelogs shared here :

26 August 2026 Update V131

  • When you touch the prompt box, it was becoming shrinked and this issue resolved

  • MiniMax H3 Upgrades

    • Continue videos using the final 1, 5, 22, 39, or 56 frames, instead of only the last frame.

    • Multi-frame continuation uses MiniMax H3’s native clip guide for stronger motion, identity, and scene continuity.

    • Replayed context frames and matching audio are automatically removed from the generated segment before merging.

    • Continuation works with both FL2VA and Ref2VA without consuming an additional reference-image slot.

    • The two Image To Video presets are now Image or Video To Video.

      • So with these presets now you can do image to video or video to video

    • Start Frame / Init Video now accepts images or videos. A video can supply its final frame or selected multi-frame context.

    • Merge videos can append the new generation directly to the uploaded source video, preserving video order and audio.

    • Folder batches retain sequential generation, per-prompt saves, continuation, and final merged outputs.

  • Now we have contextual continue feature with Continue From Last Video Frames option we have

  • This feature implemented into SwarmUI as well so read it too

  • To update : get latest zip file, extract and overwrite all, then run Windows_Install_Or_Update_ComfyUI.bat and also Windows_Custom_Nodes_Bundles_Installer.bat (bundle 100) - all necessary custom nodes and extensions automatically installed and many of them managed by us for extra security

22 August 2026 Update V129

  • Python 3.10 not anymore supported because onnxruntime-gpu 1.27+ CUDA 13 build requires Python 3.11 or newer

    • The installer will auto delete your existing Python 3.10 venv and auto generate new Python 3.12 venv if you have Python 3.12

    • Please have Python 3.12 and use it

  • This is a very big update : 8 new official LTX 2.5 presets, MiniMax H3 Init Audio (make the video follow any soundtrack with lipsync), live token meter, Face Inpainting upgrades, automatic FL2VA / Ref2VA routing, a speed update and a big installer overhaul so please read all

  • 8 new LTX 2.5 presets added

  • The 2-stage presets generate at 960x544 with 8 distilled steps, then apply the official LTX 2.5 latent spatial upscaler x2 and run 3 more refinement steps on a 1920x1088 canvas, and finally center crop to exact 1920x1080 - all with synchronized audio at 24 FPS and about 5 seconds by default

  • The Fast preset stays single stage at 960x544 for quick tests, and the First Last Frame preset is the official single-stage 1280x720 workflow : load a first and a last image and it generates the connected shots in between

  • The Expert preset exposes every stage on the canvas (Preprocess, Generate stage 1, Upscale and re-sampler stage 2, tiled Decode) with notes explaining each setting - the Text To Video and Image To Video presets are the compact subgraph versions, enter the subgraph when you want to fine tune

  • Audio To Video : load any audio, trim it with audio start and duration, and the video is generated to follow it (lipsync, action timing) - the encoded audio tokens stay frozen in both stages and your original trimmed audio is muxed into the output untouched, never decoded from latents - an optional start image is supported

  • Motion Track Control : drag points on the LTX Sparse Track Editor to sketch how things should move on your start image, and the motion track IC-LoRA animates exactly along those paths - a tracks preview video is saved next to the result so you can check the motion before committing to a long run

  • Text To Audio : audio only generation (sound effects, ambience, music cues) with the same LTX 2.5 model - no video VAE and no upscaler needed in this flow

  • Prompt enhancement (the official Gemma 4 E2B INT8 ConvRot enhancer) is off by default in every preset - enable it to expand short prompts into rich cinematic ones - and entering an LTX API key switches text encoding to the API to save VRAM

  • Models : one bundle for everything - open the latest model downloader and grab the LTX 2.5 Video Core Bundle : it now contains every model referenced by all ten of these LTX 2.5 presets : https://www.patreon.com/posts/114517862

  • LTX2.5 IC LoRA Pixel Spatial Upscaler presets fixed (FAST Distilled and Slower HQ) : the 4 helper nodes they need (LTX-2.5 Upscale Controls, Distilled Sigma Schedule, Prepare Video Chunks, Merge Video Chunks) are now part of our auto installed and managed by us extensions

  • Both upscaler presets also switched from the 24.46 GB BF16 text encoder to the 14.32 GB Gemma 4 12B INT8 ConvRot v2 one, so all ten LTX 2.5 presets now share the exact same text encoder and prompt enhancer - that is 10 GB less to download and one less model kept in memory when you switch between presets

  • New LTX2.5_Enchance_Prompt_Feed_For_LLMs.txt inside the Prompt_Generate_MiniMax_LTX_and_Demo_Materials folder : an official-aligned LTX 2.5 prompting guide for LLMs - give it to ChatGPT / Claude / Gemini, tell it what you want, and it writes properly formatted LTX 2.5 prompts for text / image / audio to video, multi-shot and IC-LoRA workflows

  • Optional Init Audio added to all six MiniMax H3 video presets (Text To Video, Image To Video, References To Video - both 8 Steps and 4x Speed) : select or upload any audio - or even a video file - in the new INIT AUDIO (OPTIONAL) node and the generated video follows that soundtrack exactly : lipsync, action timing, ambience - and keeps it as the exact final audio track

  • Leave it at (none - disabled) and the presets work exactly as before

  • Works with text only prompts, start frame and gallery references - by default the video duration automatically follows the audio length or set duration_mode to keep workflow duration

  • Init audio accepts video files with a visual trim panel - drag the handles to pick exactly which part of the soundtrack to use, nothing is re-encoded

  • Folder batches : put a same-name audio / video file next to each batch item and it is automatically used as that item's init audio (SAME-NAME INIT MEDIA)

  • Live MiniMax H3 token meter added to the reference gallery : prompt + references + resolution + duration are counted in real time against the 109k token budget so you see the usage before you queue - the meter also follows your init audio selection and trim window

  • Automatic FL2VA / Ref2VA routing : the MiniMax H3 video presets now pick the right checkpoint per prompt by themselves - single runs that have references and every folder batch item go through the Auto FL2VA / Ref2VA pipeline while plain text prompts keep the normal path - and the References To Video presets got an optional INIT IMAGE node : FL2VA uses it as the exact first frame, Ref2VA adds it as the starting frame picture reference (folder batch items ignore it and keep using their same-name init media and last frame continuation)

  • Face Inpainting upgrades : new faces to inpaint widget selects which faces get refined : 1 = biggest face (default), 2 = second biggest, 1,3 or all - faces are ranked biggest to smallest and each selected face is refined in its own pass and tagged in the tracking preview

  • New hallucination guard : every regenerated face crop is compared against its input and frames where H3 painted a neighbouring or different face over your subject are not pasted

  • Geometry lock is further improved and ON by default - removes the per-frame face shaking / tilting that the face pass could introduce

  • MiniMax H3 Text With References To Image preset also received the new face inpaint subgraph and a randomized seed

  • Speed : MiniMax H3 Sol-Attn sparse attention updated for SM89 (RTX 4000 series) and CUTLASS 4.7 in our maintained ComfyUI-TeaCache fork

  • Example on RTX 5090 : 1344x768 10 second video dropped from 450 seconds to 401 seconds with the updated speed optimizer

  • FLUX 2 Klein presets switched to my new INT8 ConvRot HQ models - almost BF16 quality but massively faster - auto download them with the latest model downloader : https://www.patreon.com/posts/114517862

  • Installer overhaul : custom nodes are now updated in-place instead of delete and re-clone so files downloaded inside node folders (for example ckpts) survive updates - a fresh clone replaces a node only when the in-place update fails

  • Leftovers of interrupted installer runs are cleaned automatically and staged copies use .disabled names so ComfyUI never imports duplicate nodes

  • Deleted venv is recreated in place without re-cloning ComfyUI - your models, outputs and user data are never touched - and the installer window now always stays open with a clear final status message

  • Python 3.10 support dropped because onnxruntime-gpu 1.27+ CUDA 13 build requires Python 3.11 or newer : menu now offers 3.11 / 3.12 (recommended) / 3.13

  • OpenCV conflicts between custom nodes fixed for good : all four OpenCV distributions are kept at one resolved version and opencv-contrib-python-headless is reinstalled last then verified - RES4LYF, controlnet_aux, Frame-Interpolation and rembg each pull different ones

  • onnxruntime-gpu upgraded to 1.27+ (CUDA runtime used by FantasyPortrait) and google-genai added to requirements

  • ComfyUI-Frame-Interpolation now installs without CuPy - it is not needed by the bundled RIFE workflows

  • RunPod and Massed Compute installers received the same in-place update and OpenCV normalization treatment

  • Windows_Run_GPU.bat renamed to Windows_Run_ComfyUI_Default.bat

  • To update : get latest zip file, extract and overwrite all, then run Windows_Install_Or_Update_ComfyUI.bat and also Windows_Custom_Nodes_Bundles_Installer.bat (bundle 100) - all necessary custom nodes and extensions automatically installed and many of them managed by us for extra security

16 August 2026 Update V122

  • Presets and custom nodes updated to improve face shaking effect that happens when automatic face inpainting activated

  • All presets now do have live preview (no increase in VRAM or generation duration) and fixed seed feature

16 August 2026 Update V120

  • I have added automatic Face Inpainting feature to all MiniMax H3 presets

  • Get latest zip file, extract and overwrite all and run installer update bat file to update

  • Also use updated extra_model_paths.yaml if you are downloading models into SwarmUI

14 August 2026 Update V117

  • Since you have requested, all MiniMax H3 presets modified and easy LoRA use added

    • If you don't select any LoRA, workflow works just fine no issues so they don't break presets and how they work, they are optional

  • 3 New presets added

    • MiniMax Music 3 Text To Music - 260813.json - New model published today

    • LTX2.5 IC LoRA Pixel Spatial Upscaler 2x - FAST Distilled 8 Steps - 260813.json

    • LTX2.5 IC LoRA Pixel Spatial Upscaler 2x - Slower HQ - 260812.json

  • Moreover, I have converted the LTX 2.5 Int8 ConvRot myself higher quality and more accurate than officially published ones

    • Someone detected one of the model is corrupt and broken - officially released ConvRot Int8 variant

    • I converted them from BF16 versions after big research and experimentation

    • So our model downloader app will download better quality Int8 ConvRot of LTX 2.5 models

  • Our Musubi Trainer convert to quant tab also now supporting LTX 2.5 Int8 ConvRot compile : https://www.patreon.com/SECourses/posts/secourses-musubi-137551634

  • LTX 2.5 video generation presets and bundles will be added soon hopefully but its quality lower than MiniMax H3

  • Get latest zip file, extract and overwrite all, then run Windows_Install_Or_Update_ComfyUI.bat and Windows_Custom_Nodes_Bundles_Installer.bat (select bundle 100) for updating all

12 August 2026 Update v116

  • To improve security, I have forked auto installed repos and bundle 100 repos and now they will be installed from secure fork

    • Only few repos left as exception which are famous : mcmonkeyprojects , kijai , Lightricks

    • I also did a massive deep security check of all auto installed and bundle 100 repos source codes and all are secure

  • MiniMax H3 presets got updated and now we have new features

  • Now when batch processing, you can dynamically set the duration of each prompt

    • It will split the file name of the prompts with _ and if last split is a valid integer, it will be used as the duration of the prompt, otherwise whatever you set will be used

    • Folder durations now use the final positive integer suffix only when the filename contains : scene7.txt → 7s; 200.txt, scene.txt, and scene_0.txt use the configured duration

    • For example scene1.txt or 1.txt will use your set duration on the workflow meanwhile scene_1.txt will be set to 1 second duration dynamically

  • Added Continue from last frame beside Merge, disabled by default

  • Continuation automatically uses Ref2VA when references exist and FL2VA otherwise

  • Updated all MiniMax presets, extension UI, documentation, and the LLM prompt feed

    • MiniMax_H3_Enchance_Prompt_Feed_For_LLMs.txt improved big time and includes latest changes and features as well

  • The most professional speed up LoRA maker LightX2V just published new speed LoRAs for the MiniMax H3

  • Our MiniMax H3 core bundles now auto downloads both FL2VA Turbo 8-step v1.0 and FL2VA Turbo 4-step v1.0 768p

  • All presets by default uses FL2VA Turbo 8-step v1.0 but you can switch it to FL2VA Turbo 4-step v1.0 768p

  • Older version 0.1 removed from downloads and presets so you can delete from LoRA folder if you downloaded

  • All presets exposes Shift value for Video and Audio seperately now

    • So you can set to 6 / 3 if you switch to FL2VA Turbo 4-step v1.0 768p - default is 12 / 3

  • Sadly LoRAs only made for non-reference model atm but works on both models until they publish specific new LoRA for reference model

  • Some people asked me if video-latent continuation input is possible or not and sadly not possible so your best option is using continue from last frame

  • Get latest zip file, extract and overwrite all, then run Windows_Install_Or_Update_ComfyUI.bat and Windows_Custom_Nodes_Bundles_Installer.bat (select bundle 100) for updating all

11 August 2026 Update v113

  • RunPod and Massed Compute startup command changed so that the running ComfyUI instance won't be accessible by public IP anymore

    • Updated files are : Massed_Compute_Instructions_READ.txt and RunPod_SimplePod_ComfyUI_Instructions_READ.txt

    • Windows doesn't have such vulnerability it is fully locally running

  • Use Cloudflared secure HTTPs to access publicly when you are running on cloud services

    • RunPod and Massed Compute installers auto installs Cloudflare

    • Follow Massed_Compute_Instructions_READ.txt and RunPod_SimplePod_ComfyUI_Instructions_READ.txt

    • Public access is optional, it will be default fully locally restricted running

  • When using unicode character containing references it was causing issues

    • For example an image with Ü Ğ characters, this issue fixed and now works

  • Just run Windows_Install_Or_Update_ComfyUI.bat to update and always get latest zip file and overwrite older files

0