Made in ComfyUI using only local models (and uncensored of course). workflow and usage is the one on this post, with reference actors that seem to work quite well or this direct link to the workflow . embedded workflows and prompts in each asset (basically the whole related output folder from comfy). Models used : LTX 2.5 for Video and 2 shots with WAN 2.2 (explossions ones, at this LTX sux) Flux Klein for images, IndexTTS for voice, Audio Ace Step 1.5 for music
LTX-2.3 is now out with better coherence both temporal and spatial. So I gave it a spin. With a hard sure-to-fail scenario. A very long continous action scene with actor and environment referencing. I know the characters change faces during the video but this is my fault as I updated the characters during the creation of the video.
In my previous posts I talked about how you can use LTX-2 as an WAN upscaler/refiner and how to add external actors and elements references without img2vid (you need an empty scene without them and need them to come into the scene). But why not both ? LTX-2 sux in action sequences and human interactions so the alternative at this point is wan 2.2 . But wan is lowres and has the same issue as ltx, no way for now to add actors in latent space. So I used the same technique as for LTX2 to add actors to wan and then reinforce them in LTX-2 using the same method. Here are some results:
Idea : Generate a very low res wan 2.2 video as reference for LTX but still pre-appending the actors and elements images at the beginning of the video,, then have the first image from the actual shot and referencing the characters from the beginning in the video. This step at 480P is very fast and good enough for characters interaction/movement coherence etc to be used as vid2vid in ltx-2. We save it at 12 fps so we can upscale with temporal upscaler in ltx. Then in the LTX step we bring the same intro images but at highest resolution possible so ltx knows how the characters actually look like in maximum detail and paints them over the lowres wan video at at a 4x resolution. So the 480p video becomes 1440p in this case (but you can go lower if you don’t have the resources, I have an 3090 and 64GB system ram). Both qwen image edit and flux klein were used for generating the actors, scene, zoom ins on the scene, removing characters etc.
I have had a lot of fun with LTX 2 but for a lot of usecases it is useless for me. for example this usecase where I could not get anything proper with LTX no matter how much I tried (mild nudity):
In fiecare zi afli cate ceva nou despre lume sau despre tine. Ieri am avut unul din acele momente de „aaaaah, deci asta era”. Nu am si nici nu am avut vreodata capacitatea de a forma imagini mentale. Prima oara cand am avut un sentiment ca ceva e in neregula a fost in copilarie cand incercam sa folosesc metoda numaratului oilor, precum mentioneaza si primul dintre articole. Pur si simplu nu reuseam sa vizualizez asta si mai mult ma frustram decat sa ma odihnesc. Mai apoi m-a deranjat atunci cand incercam sa desenez si ma loveam de imposibilitatea amintirii unui chip uman. Imi puteam aduce aminte trasaturi, proprietati si sentimente si astea erau mai mult adjective imaginare. Nimic consistent si nimic precum o imagine. Acest lucru se manifesta si acum dar l-am ignorat crezand ca asa trebuie sa stea lucrurile. Nu conteaza daca e vorba de parinti, prieteni buni sau locuri, ceva ce aduce macar a imagine lipseste cu desavarsire . Stiu ca pare ciudat ca cineva a carui principal domeniu este vizual insa dupa cum am aflat nu e chiar asa neobisnuit.
Cand eram mic de multe ori stateam intins pe acoperisul casei la bunici si ma holbam ore in sir la stele. Eram fascinat de imenistatea universului, de misterul care se ascunde in spatele covorului nesfarsit de puncte mici si stralucitoare. Mi-am pastrat fascinatia respectiva si e unul din motivele pentru care viata mea a decurs cum a decurs.
Cand m-am facut mai mare mi-am luat prima luneta, era o chestie de 3 milioane din pavilionul H, lentila frontala 50mm, nimic extraordinar dar a fost prima oara cand am distins cu ochiul liber silueta lui Saturn si mi-am satisfacut o mare curiozitate din copilarie. Mai tarziu m-am apucat de fotografie, am lasat luneta deoparte si am inceput sa explorez lumea prin prisma obiectivului. Apoi am trecut prin perioada psihedelicelor si am explorat universul prin perspectiva universului insusi. Apoi mi-am luat si un microscop si am explorat lumea sub-milimetrica.Read More
Note: This document is written one year ago and is not 100% up to date. A modified (and more technical) version was presented at GDC 2014 by Romain Sididris and it is the main technique used in Modern Combat 5. For those of you with GDC Vault access the presentation is here .
At the time it was I think the only game that used SSAO on a tablet or phone.Read More
I usually don’t write about something else than photography, but this time I feel is somehow my duty. As you know I work in the gaming industry. As an industry, it has little margin for something that is unsure to sell, something experimental, something more art oriented rather than existing bad recipes. So basically “the industry” only manufactures the same existing games over and over, with the same compromises made for accessibility and mass-market, the same action packed brainless, soulless, kitschy pieces of shit. Of course there are exceptions even in the high budget area with some developers taking huge risks to push onto the market games that have potential not to sell. Sometimes they make it, most of the time they don’t. And it’s a real pity since gaming platforms have the potential to push art to a new evolutionary step through interactivity, where the viewer becomes a part of the art piece. And that can be the one and only final step in modern arts, it cannot go beyond total immersion, it can only evolve around this idea along with technological evolution.
