wanx-troopers

LTX 2.3

LTX 2.3 uses Gemma 3 12B as multi-modal text encoder. Gemma is by Google. It might be advisable to set width and height as multiples of 32 (128 was suggested to fix some sort of issues). Frame count native to LTX 2.3 is 1 + 8 * N since 1st frame is encoded separately and subsequent ones have one latent span 8 frames in time dimension.

See also:

Gemma

refusal happens on the decoding, when encoding prompts the model has no choices to make, it’s a single pass so the abliteration only helps when generating text, which doesn’t happen when you encode a prompt … and the potential downside of using abliterated model for prompt encoding is that the model wasn’t trained with that, it may work, and change result, but overall technically it would be worse

Q: abliterated lora?
A: the point of the lora was that you can load normal model and use it for prompt encoding, and apply the lora only for prompt enhancer on prompt enhancer it’s almost mandatory considering how easily Gemma refuses even relatively safe prompts

HDR LoRa

Lightbricks have released HF:Lightricks/LTX-2.3-22b-IC-LoRA-HDR. This IC LoRa takes a normal REC709 video as guide and produces a matching video encoded using Arri LogC3 tonal curve adding details in shadows and highlihts as needed. The nature of diffusion models is that out put is always in floating point. The catch is that normally extracting more than 8bits of useful data from it is not possible because that is the models were trained on. Not in this case. It is possible to extract more bits of data from a video generated with this LoRa.

KJNodes now have HDR Preview KJ node with an exposure slider.

Supported output formats for this kind of video are LogC3 EXR, LogC3 ProRes 4444 and linear Linear EXR.

LTXVHDR Decode Postprocess node is provided for saving such videos in LTXV custom nodes. export OPENCV_IO_ENABLE_OPENEXR=1 environment variable may need to be set outside of Comfy. Kijai: “I think my Marigold nodepack hijacks and sets the env var before other imports, or at least used to, I remember dealing with that issue”.

Q: If I just save the model output directly, which should be in LogC3, then import to Resolve, set the input to be ARRI C3 (or whatever it was called), it seems to work fine, is there a reason to convert it to linear before export? Didn’t really see any visible differences doing it via EXR vs ProRes4444 myself either

Nathan Shipley:

A: Prores 4444 is 12 bits per channel, the linear EXR should be 16 bit float; you’ll probably only notice the differences between 12 vs 16 bit in more extreme grading, or in potential banding that might show up

Richard Servello:

the raw output is logC tho so even at 10-bit it will have a similar range; it seems to also fix noisy edges and clean up some blurry footage could be interesting. the curve is wrong for their linear conversion

Nathan Shipley:

LTX HDR IC-LoRA turned up to 1.5 Distill Lora at 0.6 euler, 8 steps, linear_quadratic CFG 1.5 can definitely help to turn the lora strength up above 1 for shots like this..

the VAE as it works right now can’t generate values outside of 0...1. this LoRA is working by generating values compressed into a LogC space - so, when it gets expanded out from LogC into a working HDR space, overbright values go above 1. (in Nuke, I set my Prores I’ve rendered to be AlexaV3LogC … hovering over the bright sun areas shows values around 18

Nathan Shipley’s wf: ns-LTX-2.3_ICLoRA_HDR_Distilled_Simple_NS_01; “I included the LTX EXR write node in there, though Prores 4444 is probably fine most of the time!”.

Richard Servello has added his nodes RS LogC3 HDR Decode (hdr_linear, raw, sdr_preview) and RS EXR Sequence Save to GH:richservo/rs-nodes. “it uses proper baked ocio luts” “now the hdr_linear will output proper color and the sdr preview is aces rec709 converted”.

Alternative similar LoRa from T2/Greg: gregt/lora_weights_step_07000

Potentially useful:

Settings to view linear EXR created with this LoRa in AfterEffects with OCIO engine from Oumoumad: ltx23-hdr-view-ae-ocio; his advice for DaVinci is to use sRGB linear setting.

Prompt Relay

Prompt relay is a new technique which sub-prompts to specific parts of the video only though attention masking.

GH:kijai/ComfyUI-PromptRelay to implement “Prompt relay” technique using prompts like:

a does this..
|
then does that

Alternatively people have been experimenting with [4-8s] ... style of prompting style of prompting and no prompt relay. [4-8s] ... seems to sometimes work and sometimes not. At one point it seemed like using ID-LoRa would make [4-8s] ... work better but that is far from certain.

[NAG] just doesn’t work with prompt relay because both patch cross attention

prompt relay has to use sdpa for crossattn because sage doesn’t support the masking

[epsilon] controls how much the mask supresses the other local prompts; higher epsilon allows the prompts to bleed into each other more, depending on your prompts it may create better flow, with low eps the cuts are hard

prompt relay only needs latent to know your input size, it can just be empty latent too
Q: …does it need to be audio + video latent, or could it just be video?
A: iirc it should just be video

Wiring help: prompt-relay.

The Shadow (NYC)’s prompt relay setup: the-shadow-prompt-relay.

SirAxe’s experimental nodes to apply LoRa-s to different parts of the video selectively, extending functionality of prompt relay: GH:kijai/ComfyUI-PromptRelay/pull/1

Context Windows With Ltx 2.3

Drozbay on context windows implementation for LTX 2.3:

Getting dialogue to be split perfectly among context windows even using split conditionings is gonna be very hard to get right. I have gotten it to work by creating a gen at low resolution with no context windows and then upscaling by 3x or 4x to get to the final resolution, using a partially masked audio to get it to increase in quality without totally losing the content. No context sliding specifically. I’m injecting the first frame into each window as a pseudo-reference. It doesn’t always end up looking that good.

From The Makers

HF:Lightricks provide

They also provide under LTX-2 umbrella

May 2026 saw addition of HF:Lightricks/LTX-2.3-22b-IC-LoRA-LipDub gated LoRa for re-dubbing vidoes - the intent was to keep the original voice; based on GH:justdubit/just-dub-it undetlying research.

Alternative Weights Packagings

Nodes Of Interest

Q: do i have to intall the LTXComfy repo or is LTX native supported?
Drozbay A: Comfyui native nodes are still missing a few features so it’s probably a good idea to have the LTXVideo nodes installed, although KJNodes also has nodes that cover most of those gaps

What is this?.. kjnodes-enhance

IC LoRa-s

IC LoRa generally stands for “in-context LoRa” a type of LoRa. In colloquial speak “IC LoRa” generally refers to one of the IC LoRa-s released alongside LTX 2.3: Union Control, Motion Track or Depth/Canny/Pose. Additional Python code to use them: GH:Lightricks/ComfyUI-LTXVideo.

One way to use IC LoRa along with a reference image: ic-lora-guide-plus-latent-guide

Drozbay:

the guide latents are all just stacked together at the end of the noise latent and assigned rope indices to indicate where they belong.

On why the above method works better than alternative of supplying the reference image as an image not directly as a latent:

I think there might be something else going on here, like the ordering of how they are placed if you use the different nodes, or maybe some default settings getting applied for one and not the other

Q: Shouldn’t providing the latent at index -1 be the same as providing an image guide at -1?
A: there’s no node to do that with image, only the latent guide node currently actually places it at negative index, the other nodes use python indexing and start from the end instead

what I’ve now been doing with LTX is putting ref latent as the extra cond AND as I2V first frame it has minimal effect in the cond (guide), but enough to kick start it so the I2V cond takes over;
Yeah, I’ve been doing similar, sometimes with lower strength on the inplace latent frame

Alisson Pereira’s custom node which adds both reference image guides and guiding video for IC LoRa-s to altents, license unclear: LTXVAddGuideWithRefs.py

Keyframing

Different ways to do FLF: LTXVAddGuide, LTXVImgToVideoInplaceKJ.

YT:What Dreams Cost/Guide to Prompting and Keyframing I2V GH:WhatDreamsCost/WhatDreamsCost-ComfyUI; an independent attempt to re-purpose the What Dreams Cost nodes to use a character sheet YT:How to use Character sheets with LTX 2.3

Zueuk:

as i understand, with guides you “never” get exactly the same frames even if you set the strength to 1, so for extending i’d probably keep masked “inplace” data but when we put frames anywhere except frame #0, guides are probably a better choice

ImgToVideoInplace

Alternative to using guides. Kijai’s version also allows to specify which frame to apply to: LTXVImgToVideoInplaceKJ

Drozbay:

The inplace nodes are very simple: they don’t add any guide latents or anything they just replace the target latent frame with the encoded image and mask it so it doesn’t change

Guides

guides are latent+mask but they exist at the end of the sequence, and are applied to the position through RoPE; so basically they are stored at the end, keyframe info is stored (through the conditioning in comfy) so it knows what frame it should affect; that’s why using full guide video is so heavy, and that’s why the newer guide method halves the guide resolution

LTXVCropGuides removes the guides from latents after generation: cropGuides.webp

Tooltip on LTXV Add Latent Guide from LTX Video node pack suggests that one may be treating index of -1 to mean “before the 1st latent” e.g. before subject enters frame.

YT:4TE-QbtkiGQ for some advice on using guides to inject frames.

YT:nekodificador reported plugging the same guide image twice via both of the following nodes increases motion twoGuidesMoreMotion; twoGuidesMoreMotion2 - “just renamed, they are regular AddGuide”; Drozbay on the difference between these two similar nodes: “Nothing should be special about the IC Lora guide version except that it allows you to control the latent downscale factor. and it also doesn’t have the crf option”.

If multiple guides are used multiple LTXVCropGuides nodes may be needed: ltx23-multiple-guides

Hashu [not entirely confident]:

If you have guides for ‘different’ frames then you need only 1 crop guide. So in your example if you change the frame idx and have one at frame 0, the other at frame 5 and the other at -1 you only need 1 crop guide node. But the moment you have an overlap in frame idx then you need 2 crop guide nodes. So if you are using an ic lora + a first frame guide then u technically have 2 guides at the first frame so you need 2 crop guides. ltx23-multiple-guides-2

Controlling The Camera

N0NSense controls camera using a schematic video or a box/room everything happens in converted to a depth map + ic union control LoRa. His a2v videos created this way look great.

Cseti’s LoRa to replicate camera motion from one video to another HF:Cseti/LTX2.3-22B_IC-LoRA-Cameraman_v1; README and workflow: HFdatasets:Cseti/ComfyUI-Workflows:ltx/2.3/ic-lora-cameraman; “This one took around 20-24 hours to train with 77 video pairs. And I also made two more runs one with 128 and another with around 40 pairs. But this one looks the best so far” “I used videos from pexels”

I2V

Hevi:

lowering guidance stregth to 0.3-0.4 for the ref image helps with the watercolor artifacts

Motion

To fix motion arfiacts ppl often generate at 50fps. Sometimes 35

N0NSens:

50fps is the only solution to reduce garbage in motion, I found. Initial higher resolution/res_2s/more steps/diff models/distill loras str… nothing helps.. [and then pick every other frame via VHS nodes and clean up with Wan]

It seems addition of VBVR LoRa might be helping make motion less smudged.

See also: Guides section for YT:nekodificador’s trick to increase amount of motion.

Inpainting

Extensions

It has been reported that LTX 2.3 extends videos quite well forward but a method of extending backwards without a LoRa hasn’t been worked out yet. YT:LTX2.3 | How to Extend presents the latent extension method using LTXV Audio Video Mask from KJNodes. GH:ckinpdx/ckinpdx_comfyui_workflows contains a latent looping workflow which again apparently uses the same technique.

Note that Sir_Axe’s HF:siraxe/MergeGreen_IC-lora_ltx2.3 can be a useful alternative.

Multi-Pass Workflows

Ppl often use workflows which start with a sampler generating at a low resolution and then (sometimes a chain of) 1.5x or 2x upscalers, more often 2x. This is said to improve motion but blur small objects with texture.

Zombiematrix:

use the either of the default templates and change the starting pass to go to .25 of the starting resolution (instead of .5)
then add a second upscale pass. thats all it is. you can use either of the upscalers (1.5 or 2.0)

Jonathan (WhatDreamsCost):

I’ve been testing a 3 stage workflow for the past month, mainly because I see better motion when the 1st stage is at a lower resolution. Also I like it for testing prompts. Since the 2nd stage is faster, you can get a clearer preview quicker when using the tiny vae compared to using just 1 or 2 stages. And when you use CacheDiT on top of that, you can test prompts even faster.

Mark DK Berry:

For LTX, I go 480 x 201 first pass so I can make structure quickly util I get what I want … then upscale twice for 2nd pass from v2v in LTX and I am at 1080p in 10 mins, and final is polisher low denoise to fix eyes and whatnot, often with a WAN model denoise 0.2 using USDU; any higher [denoise] the tiling or weird speckldeing creeps in but you could do a full detailer method but the USDU is faster and works on low VRAM. I am using 1.3b for this atm. still testing it, but its half the time of the 14B for me. … just tested Wan 2.1 self forcing DMD 1.3b and its good enough for most fixups at 0.2 denoise … 241 frames at 24fps

Q: 40 steps on base dev at 0.25 res, 3 with distill at x2 upscale, and 3 again with x2 upscale? Is this all linear quadratic schedule?
A by Garbus: That’s it, and er_sde across all three. Distill LoRA on 2 and 3 set to about 0.6 usually is best. (I only use the preview node on the first sampler since it can cause an OOM on the upscale, and there’s nothing new to see there anyway.)

Q: When you run … Wan step how well does it deal with identity preservation?
A by Mark: use phantom 1.3b or WAN 2.2 VACE self forcing 1.3b (which is i2v effectively). and keep denoise low but with phantom you cant go over 0.5 it gets weird but consistency is kept so far in tests both with phantom and VACE 1.3b. HuMO would be better but my rig cant wait for it. I wish MAGREF did a 1.3b I would have used that but. I tested 5b wan but it wasnt very good which was weird.
Q: Do you end up doing context windows for a long base LTX gen to not run out of RAM? A by Mark: no, I use USDU then you dont need all that. it handles it. its in the wf links for my video pipeline. have a look.

I am not sure if HuMO will work with USDU properly. I currently use either Phantom 1.3b or WAN 2.2 VACE 1.3b only because lowVRAM else I would use 14B and I use it in two stages, once to detail the 480x201 intiial video then again with USDU for the final polish at 1080p after its been through LTX upscalers.

previously was two x2 upscalers in series but today added in a 2nd sampler to the first upscaler and its way better. I still have some tweaks to make though but also introduced the VBVR lora and the 1.1 distill came out so it might also be that helping improve thing… the guts of it is in my video pipeline MBEDIT-v2v_LTX23_Upscaler_DevQ5KM-Audio-In_1080p_vrs7.json https://markdkberry.com/workflows/research-2026/#video-pipeline-workflows

I was using phnatom 1.3b to detail my i2v first stage 480 x 201 that I did that size to get structure quick as I could changing prompt til I got there. it was fast relatively. but the phantom did a nice fixup. then into the upscaler wf with ref image. but it had its faults. today introducing x2 samplers to that upscaler stage instead of x1 sampler with x2 upscalers has meant I dont need the phantom stage, but I might keep it in anyway.

when using WAN models with a sampler as a detailer anything over 0.78 I think completely changes the shot to the prompt. polishing was always down around 0.1 and stronger fixes up around 0.2 to 0.4. (USDU it goes more to 0.5 or even higher in some cases …)

Mark’s video: YT:Video Workflow Pipeline (April 2026)

N0NSens on flickering in three pass workflows:

I think it’s because of LTXVImgToVideoInplace at upscaler passes. At first frames it’s combining lowres latent from prev pass and your input image. So far, the only way I found to reduce this effect is to cut off first 3 frames and lower Inplace str to 0.7

just matching to ref helps in this case

“native SUPIR implementation PR” from Kijai includes a new color matching node

N0NSense

In my experience, the higher the initial resolution, the better the result. Therefore, 2 passes produce a better result than 3 (especially noticeable on small patterned textures).

David Show

but the higher the first pass resolution

[audio] “generate … once … all subsequent upscale passes … reuse the originally generated audio rather than reprocessing it”

first pass with a modality scale of 2.5

It was reported an overly high sqare resolution (1920x1920) after upscaler can cause color shift issues. Wider or smaller resolution like 1536x1536 reportedly can help.

official-video_ltx2_3_i2v.json demonstrates often used up/downscaling tricks in multi-pass pipelines.

Wan-Various_Refine-Upscale could potentially be used to enhance LTX videos with various WAN models.

why … 4 step sigmas like 1, 0.85, 0.7250, 0.4219, 0.0 on the last upscaler sampler when 6 steps is clearly superior in results?

Richard Servello

Ltx is meant to be 2 stage but a lot of people insist on one pass and complain about artifacts

Mark DK Berry:

One thing I love about LTX that WAN is [bad] at, is I can shove the same video through it over again and no problems. WAN will blister the contrast to hell if you try that, VACE too.

Training

On character LoRa training: “just 30 images with 10 repeats and 10 epochs, so quick and dirty - AkaneTendo25 fork of musubi-tuner-ltx-2 - success”

10 videos from 2-8 seconds each … if overfitted reduce strength of LoRa

Training IC LoRa requires twice the VRAM and twice the time compared to traditional LoRa-s. 5090 should generally be capable of.

Oumoumad:

I never needed to go beyond 5000 steps, in fact most of the time even in 1500 steps you already see your desired effect

mamad8:

Using split sigmas with the distill Lora (strength 0.5) to set cfg 2 for the first 2 steps and cfg 1 for the remaining 6 steps helps A LOT, especially audio but also overall coherence and motion

mamad8:

If you have trained a character Lora but you’re often using it with another (or multiple) loras you might have seen that the character is less accurate (it depends on how much the loras you use have been trained on specific faces) : select your best 4-6 close ups of your character and train a new Lora rank 4 on top of the models you use (and your previous trained Lora) for 500-1000 steps, the result will be absolutely night and day. Currently I don’t think any trainer allows training loras on top of others (it’s not finetuning, simply merging the provided loras to the base model before training a completely new one. This way the loras already trained do not lose anything they have learned and they overfit a lot less). I’ll release the training code to train on top of loras soon (fork of aitoolkit)

GH:richservo/rs-nodes LoRa training inside ComfyUI “adding ffn chunking fixed the OOM” “I’m … training on 576x576x97 … it’s much more efficient … no idea why” “uses ollama server”

it’s only 2 nodes. Data Prepper and LTX train; if you have less than 32GB VRAM you want to plug in to the model vae and clip; vision is if you want to target non-human objects; if you want to always include people in your output then select face detection; if you have specific people you can plug on into the pin, or specify a folder of named images; if you are doing audio, provide voice samples of 15-30 seconds for captioning; you can even have it name locations if you add items to a folder and point it to it

GH:vrgamegirl19:

so when training with images you end up loosing motion the higher the steps … I took my 1000 step lora which keeps motion but does still add the style a bit and used that on the first pass then put my 5000 step lora on the 2nd pass … add the that VBVR to the fist pass only

first video has the 5000 step lora on both first and 2nd and the 2nd video has the 1000 step on first and 5000 step on 2nd pass.

burgstall 2026.04.22

AFAIK the LTX official trainer still is the only method for IC-LoRA training. At least thats what everybody I know uses

crinklypaper

I think musubi fork can do ic-lora training, I havent tried it yet though GH:AkaneTendo25/musubi-tuner:ltx-2/docs/ltx_2.md#ic-lora–video-to-video-training

Ada:

Yea. Most importantly that has CREPA. Which is super important for video training and nothing else has it yet crepavideo.github.io Never train video without it. Actually simple tuner also has it and does a better “similarity” threshold to shut it off which works better. I ported that over in my own fork

I use a similarity threshold of 0.9 the same way simple tuner does and shut it off after that

burgstall on training bodypositive LoRa:

140 video pairs

[Alisson Pereira] Q: So you created the first frame and then used a control union to assemble the dataset with the first frame following the original video. A: yep, exactly

Oumoumad:

personally I’d take it in reverse, ie just find fat people videos, and turn them skinny with image model, the gen the rest with control. And use the original fat ppl.videos as target. Always ensure target is as natural/original as possible

burgstall:

I just started 512x384x25f pairs and its at over 31Gb

had to lower rank to 32 to fit

Sound

The role of a “solid mask” is not clear to the writer of this page but apparently it may be used before feeding Audio latent into sampler: solidMask.webp

Some ambience noise on input audio may make ltx generatevlipsync better.

Mark DK Berry:

I use vibevoice multi speakers and then shove 10 second (about 3 lines of dialogue max) in to LTX … LTX 2.3 … “add subtle ambient noise” trick

It would be natural when using an IC LoRa to supply reference image via LTX Add Latent Guide with index=-1 completely separate from the the guide video. However it looks like existing LoRa trainers don’t yet support this. This forces training of IC LoRa-s that take reference creatively - as 1st frame of guide video, on the side of guide video etc.

GH:vrgamegirl19’s training node for quick training of LTX 2.3 single character LoRa-s: YT:9Z_glyAHE1k.

Manual Sigmas

David Show: 1.0, 0.995, 0.99, 0.9875, 0.975, 0.65, 0.28, 0.07, 0.0, “it’s a three stage workflow, but those new sigmas are just being used on the first pass.”

Reddit post.

V2V

V2V can be done either via IC Union LoRa-s or via latent denoise. Unmerged Context Windows PR to make V2V simpler: GH:Comfy-Org/ComfyUI/pull/13325.

Hicho:

t2v latent low denoise from another video

Basics

Node Packs

ID LoRa

ID-LoRa apparently provides both visual and audio character likeness. Additionally it seems to support timed prompting otherwise not available on vanilla LTX 2.3. Because one of the checkpoints is called “talkvid” ID-LoRa is also referred to as “talkvid”.

TalkVid flavor possibly supports the following style of prompting - though it remains uncertain

[abc] A young …

[scene] …

[sound] soft…

[0-1s] [abc] raises …
|
[1-6s] …

Gleb Tretyak:

not 10/10 follow. timings should be sufficiently calculated by user, otherwise it’s ignored by the model I believe; and prompt relay makes it worse

Draken:

id lora is just for audio really; it helps in the sense that the lora itself might do better at perversing the ID in i2v mode though

LoRa-s, Alisson

LoRa-s

Less Verified LoRa-s

Non LTX 2.3 LoRa-s Relevant to LTX 2.3

Workflows and Workflow Collections

Joke And Experimental LoRa-s

This site is open source. Improve this page.