Model Arena - https://app.fuser.studio/view/b7a2a296-6121-45f7-bd0e-dfe1ae54bc32 For the community a tool to help you on the last day with video, the tedious type of testing setup to build for the longterm. 🏃‍♂️For your fuser.studio project are you in a rush but still need to pick video models for your ideas? 😵 Or are you not sure how to go about comparison and the organization of everything without ending up with a confusing mess? I made the Model Arena template just for you to speed UP shortlisting generative video models text-to-video comparison on fuser.studio. 🐐 Goat if this helps you or you think i should make other version so of this for image-to-video vid-to-vid etc

The brand concept In the template for the #fusercocreate hackathon is a teddy bear brand working on lookdev for adspots, commercials, marketing videos and potential tv series DEFENDERS OF IMAGINATION. (on contra I'll provide ~3 sample vids)

What a Model Arena does, It breaks down fuser.studio's available text-to-video models across three categories based on cost and speed put right on the canvas for quick iteration. Then a creative-code node to make a video review gallery with categorization and yay or nay voting for stakeholders (vibe coding in progress 👓 🏖️ ). With a compositor to bring a final three videos together with a text overlay of the prompt or a narrative you add (NOTE , this is WIP as the compositor is not working from an error for me "Video is not supported in compositor on this device." prolly cuz I work on a potato 🥔)

The questions I used for the categories are: does it run in under minute or does it cost under 600 credits.

Cheapest v. Fastest near a minute, Mid-Range cost and ~two minutes, Expensive or Slow. Node labels includes a cost(💸) & speed(⚡) info on each node for convenient reference.

Examples, Fast SVD Video Generator 💸69 ⚡92s which is cheap and fast but subpar results with a quicky prompt. Meanwhile there's Pika 💸276+ ⚡237 which is cheap but it is slooow so it goes in the mid-range.

So to run a test on the cheapest category is ~2260 credits if you run EVERYTHING.

But how to craft prompts to test video models? There's a lot of writing about prompt engineering. For me one of my approaches for testing generative models is INCOHERENT storytelling. Things that could make sense to a person but whose keywords and context are confusing to a machine. The example prompt I used in the Model Arena template is basically this: teddy bears at a picnic defending against monsters under the bed that is a crayon drawing on a fridge.

So any model has to convey how a picnic is happening but also there to be a bed in the scene. Will it put a bed in a park, or a picnic in a bedroom, where does the fridge exist and how does it appear, will the crayon drawing be applied inconsistently or not. I also set it up with heavy negative-prompt.

How I went through building, configuring and organizing this I used fuser.studios add nodes ui(shift+A) to search and filter to the text-to-video models then favorited them so I wouldn't have to repeat searching & filtering for dozens of models. Then going down that favorites list I broke out the costs and speed. 📝One issue is some "text-to-video" models are NOT purely text only capable, and require images and or video; those nodes I put in their own groups above each category and set the image/video properties to be exposed.

For configuration, I set it to hide common excess properties while exposing other important properties shared by most models , or unique capabilities for some nodes. Common properties shown list (non-exhaustive): Prompt, Negative Prompt , Duration, Seed, Model , Expand/Enhance Prompt (disabled), Turbo/Fast mode(disabled), LOOP, NSFW filter(enabled).

For consistent aspect ratio I set them all to 16:9 and hid that input. I also disabled any Turbo/Fast or Expand/Enhance Prompt settings so there's more initial control. I enabled the NSFW filter when available to help prevent surprises while at work 😲.

Additionally for TURBO/FAST & LOOP capabilities I broke out separate copy nodes for comparison to normal model nodes wiring them to a parent. 👀 Notice how on some by using counter-wiring for Turbo & Loop I made a soft conditional logic for child nodes capabilities to still be controlled by that model-nodes primary parent it's a copy of.

On some models I've exposed their special characteristics such as motion.

Canvas organization I used groups for the three categories, and the invalid nodes requiring media input. Used a PROMPT node to show the text list of models which is formatted with a three dashes delimiter(---) that lets the list be piped to a TEXT-SPLITTER(pattern) that itself sends it splits to a text-splitter in each category group. This makes it so there is a category specific list on each group for faster cross referencing but editable in one place. So if you decide to edit/add/remove nodes this minimizes the amount of places to edit. (you'll still need to keep labels of individual nodes & groups updated, no way to automate that so far afaik)

Really impressive!