V4 - Optimizations and ease of use
- Brought back partial daisy chaining so the flow moves all the way through the flow
- Saved latents only needed when extending further than 2 batches.
- Proper explanation notes on the 4 groups use and chain.
- added appending to upscale to hold seams properly.
- Added Beta57 8,6,4 step sigma values as options
- Added main switches to easily turn things on/off
- fixed small bugs.
v3.1 - fixes --- really getting sick of breaking my own stuff fixing stuff...
- Fixed preview audio for base video extending. Pretty sure the rest of the base generation worked flawlessly.
- Fixed upscale latent overlap
-- this was my fault for pushing out a fix for another issue before testing (sorry)
-- Full issue was overlap latents were not inserted so seams were not carried over causing clear changes on batch change to elements in the video when upscaling
-- This should be seamless now with testing done on multiple batch counts and all 8 batches tested.
- Few small interface changes
- The first upscale extender has (MTB nodes) audio duration node they are not needed for the flow to work, i forgot to remove testing stuff from only the first extend group for upscale. The audio cut can be done differently and i was experimenting. You can just remove the MTB nodes in the subgraph and you shouldnt need that set for the rest.
Full minimax V3 - changes
- Changed to full latent batching. Removed all audio timestamps.
- Frame perfect audio both generated and loaded
- Latent only passthroughs for easier chaining
- Added an actual interface with switches for everything
- Cleaned up inputs and outputs for easier chaining.
This should now be much easier to use.
Bugs to be fixed eventually
Preview of the main renders for the extend groups audio is messed up
To fix inside the extend groups subtgraph is a trim node near the output. The division needs to be swapped. a/b not b/a
Fix to extending upscale that i dont know if is in the live version that mixes up the overlap
Full V2 -
Fixes
- Audio masks for generated audio now work to extend.
- Bypasses now work properly
- Previews now show only one batch
-- Enabling any previews forces the flow to decode a segment. It can slow the process.
-- Because we are no longer using the video output of the subgraphs It never needs to go out of latent space. Enabling the previews forces the flow to decode the segment.
- Latents are now decoded all as one.
-- (thx obvom) Batched vae decodes cause temporal issues on the seams.
-- Workflow now uses the full latent video for decoding, and encoding.
I learned so much.. i think my vace flows could do much of this.
Full
This is ref to video with extending and upscaling
- Ref to video has
-- 6 Images, 3 audio ref inputs.
-- 3 batch groups that can be daisy chained further if needed
-- Custom batch counts with syncronised audio either generated or loaded.
-- Loras, prompts, seeds, refs can all be changed per batch.
Upscaling
- This could be its own flow, it does not need the base generation, you can load any video or even saved latent in the upscale section if you want.
-- 3 upscale options, RTX, latent, or resize
-- Loras, prompts, seeds, refs can all be changed per batch.
-- Auto batching
Vid 2 vid
This is an addition to my extending flow
Many of the options are the same.
This includes a video loader and allows you to batch copy videos
Includes memory insertion
This is a beta because, i cant test everything with it. If it breaks, let me know i will try to fix it.
All three extending flows now are in this version with memories options.
Extending
I made this originally to prove to myself that latent extending still contrast shifts. But i never can just stop there.
Latent extending can work well, but its not perfect.
Also this works off frames... not time. I got a headache with the audio/timing problems that came with the math needed for working with time. Minimax needs certain frame counts, allowing you to input seconds was useless because i had to convert all times to frame counts anyway. 10 seconds is not 10 seconds for minimax... its always a bit more or cuts and its less. This flow skips that and just asks for the frames. It will just limit your frames and overlaps to what minimax can do properly. No asking for 5 seconds and getting 6 or 4...
This comes with 2 flows
1 generates the audio with minimax
1 loads an audio file you already have to sync, music, talking etc.
I want to give a shoutout to obvpm for his nodes and flows
This guy gets it, its nice to see someone trying to make hard things easier :)
https://github.com/chanon/comfyui-obvpm-timeline/
These thoughts are basicly mine... he knows what i know about this extending stuff. I really like this guy. Give him a sub, like, comment. He deserves it.
Upscaling v2.
CORRECTION - The frames in last batch number is wrong its telling how many are missing from the last batch. Keep this as low as you can. You can leave it like that just know what it means, or you can fix it with.
To fix create math node a = batchcount b = the old calculation int. this will give the last batch count.
This was to help people with lesser systems.
Load a video you already have, add refs, prompts and upscale in batches.
Two ways to use this
Batched like default
daisy chained to do it all in one go.
Batched should allow anyone to upscale a video.
Daisy chain is for people who have the system to do it all.
Minimax can make 30+ second 0.2-0.4 MP videos, This will upscale it to whatever size you can run for the batch size you select.
Added a fix to the final size. It should now output the same length video you put in.
Removed all base video generations this is a bring your own video flow.
Upscaling
Uploaded a fix and reworked a few things to explain better how to start.
This can help remove contrast drift from extending videos somewhat.
This was made to upscale minimax videos in batches to allow smaller longer videos be upscaled to higher res.
Can make a video or load a video you already have.
V3
Advanced minimax
- generate videos from ref images.
- advanced extend videos
- use custom audio
includes a beta flow for face fixing that allows batching.
Minimax face fixing (inpaint)
This is my old vace workflow edited to use minimax instead.
Old vace flow - https://civitai.red/models/2570937/wan-vace-21-flow-archive?modelVersionId=2894367
- Cuts out faces to inpaint.
- Can keep/change/create audio
Enjoy :)
V2
Simple minimax
- generate videos from ref images.
- Simple Extend videos
- Beta memories for extending
- use custom audio
I dont think i explain everything well enough, if any questions ask in comments i will add any answers to the next flow.
Best ref2va distilled lora
https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main
MiniMax-H3-Ref2VA-Acc-8Step_pruned_comfy.safetensors
Strength 0.4-1.0
I debated releasing this as its not much more than the basic comfy workflow
Why use minimax
- Great prompt listening
- Combine multiple subjects easily.
- Quality lip sync and voice clone
- 5-15 second video clips 300 frames at 1MP.
- You want to create a story.
This uses A LOT of resources, crashed a few times, soft locked too when i went too big, but 1MP is not a tiny video when i ran about the same for LTX and much smaller for vace. Its slow, but its doing more than most.
I think it can extend like vace could, and without a contrast shift. Its down to learning the prompting to extend from the last 8 frames of a video.
Minimax is a beast for storytellers.
(fyi my gpu is stuck on PCI3 atm due to my cpu)
From clean load of comfy loading models and render of 1MP 311 frames at 8 steps took with my setup 23 min 21 seconds 158s/i
Output is 1376x768 at 12.75 seconds