LTX-2.3 Dev - bad quality
Hi RuneXX! Thank you for all you are doing for community and your excellent workflows! Experimenting with different LTX-2.3 models I found that Dev model has a big potential. It is better than distilled in terms of visual effects quality and motion. I mean Dev model without distilled lora at the 1st gereration stage. With previous model LTX-2 Dev I could get very nice generation with 30 steps and CFG 4. But LTX-2.3 with the same approach gives bad results, looks like lack of details, but motion is much better than distilled. I tried even 50 steps but all the same. With Distilled lora at 1st stage and cfg 1 it works well, but it doesn't work well without it. Do you have any idea how to set dev model correctly for good quality?
I use "ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors" model and "ltx-2.3-22b-distilled-lora-384.safetensors" lora. Try it on your basic I2V T2V work, just reconnect model node directly to CFGGuider and connect LTXVScheduler node instead of ManualSigmas. Then set CFG 4 and 30-50 steps. At least with previous LTX-2 it works well.
Here is I2V Dev CGF 4 and 50 steps - you can see much more motion here, but image is gradually falling apart
Here is I2V Dev + Distilled Lora CFG 1 and 12 steps - good image, but lack of motion and visual effect
As I understood CFG greatly influences the dynamics with Dev model. Distilled with CFG 1 can't generate such action. It remains to be seen how to improve the quality. Trying to work with sigmas as well because it make sense to me
Did you try with the LTX multimodal guiders? (that does a separate CFG for audio and video)
And what sampler (scheduler) did you use?
Not yet. I use euler ancestral cfg pp on a 1st pass and euler cfg pp on a 2nd pass
I found that euler ancestral cfg pp not working well with dev model, euler or euler ancestral are much better. It removes oversharping with CFG >1. But the problem with the lack of details remains even with >30 steps
I agree, euler seems be work better than euler_ancestral (less color shifting too). At my tests DEV model seems be follows prompt little better.
Euler is usually a safe choice. Will try some comparisons. The cfg pp stems from LTX themselves that used them alot.
But maybe euler could be the default one in the workflow, and let users try others themselves (res_2s, cfg pp , lcm etc). Since euler is usually always giving decent results
There is a lot of sampler discussion around the web with the dev model, and for sure no consensus to what works best.
Using Google AI to summarize the user experiences, gives this:
Top Recommended Sampler Combinations
Best Overall (Official Recommendation): euler_a_cfg_pp + euler_cfg_pp.
Best for Detail/Texture: res_2s (Stage 1) + lcm (Stage 2).
Note: Using res_2s in both stages can lead to over-cooked, high-contrast results.
Best for Smoothness: euler_a_cfg_pp + lcm.
Fastest Workflow: lcm in both stages.
Alternative: euler + euler (good).
Key Settings for LTX-2.3 Dev (Two-Stage Pipeline):
Stage 1 (Base Generation): res_2s / euler_a_cfg_pp with 15–20 steps (if using distilled lora at low value at stage 1)
Stage 2 (Upsampler/Refiner): Uses lcm or euler_cfg_pp (3–5 steps).
Steps: 30–50 steps are generally needed for the pure dev model to ensure clean results
CFG: Low CFG is better; start with ~3.0 and drop to 1.0–2.0 for higher stability.
And that seems pretty accurate to my own experiences as well ;-) (except for the LCM being fastest, I think euler is even faster.. at least my end)
In process of adding more DEV workflows, and been testing a bit what works best.. .
But euler/euler is a good combo and a safe choice
And then you have the whole rabbit hole of linear quadratic vs exponential vs sgm_uniform etc.. its a ton of possible variations, but most of them are quite similar ;-)
Here is I2V Dev CGF 4 and 50 steps - you can see much more motion here, but image is gradually falling apart
What if you try lower CFG?.. even as low as 2
a little test run with euler_a + euler ... dev custom audio workflow