Docs / Guides
Images and videos from text
Generate pictures and short videos on your own workers with open models from Hugging Face, watch them arrive live, keep the good ones and save them as a dataset.
Two blocks under AI turn text into media on your workers: Text to image makes pictures and Text to video makes clips of up to 10 seconds, some with their sound. Both download an open model from Hugging Face the first time and keep it in the worker's EsaSim cache. Your prompts and results stay on your workers. Use them for synthetic training data, test material for detectors, or ideas for a scene.
What you need
| Text to image | Text to video | |
|---|---|---|
| Where it runs | An NVIDIA GPU, a Tenstorrent card or the CPU | An NVIDIA GPU. The CPU works only for a test: a video takes hours. |
| Default model | Vega on a GPU, Vega RT (0.7B) on a CPU | Wan 2.2 TI2V (5B, 720p): best on a 24 GB GPU, slowly on 8 GB |
| Disk for models | About 3.3 GB for Vega, 6.9 GB for SDXL, 34 GB for FLUX.1 schnell | From about 20 GB (Wan 2.1 1.3B, with its text encoder) to about 70 GB (LTX-2.3, LTX-2.5) and more than 100 GB (Wan 2.2 A14B) |
| Hugging Face token | Optional; required for gated models such as FLUX.1 schnell and LTX-2.5. Save it once in Profile > Keys. Every block that takes it uses it when you run a flow. | |
The worker also needs Run code from projects turned on (Computing > your worker > Worker settings). The first run installs the Python packages the block needs, once per worker, and keeps the worker's own PyTorch.
Make images
- In a project's Blueprint, add Text to image (under AI) and an Images view. Connect the block's live telemetry and its Images output to the view.
- Write the Prompt: the subject, the place, the light and the style. Set how many Images you want (10 by default, up to 100,000).
- Optionally add Variations, one per line (at night, as a watercolour). They are added to the prompt in turn. To use many prompts, connect a table with a
promptcolumn or a text file with one prompt per line. - Press Run. Each picture appears on the View page as soon as it is made.
Each image uses its own seed: image n uses First seed + n − 1, so the same seed and prompt give the same image again. Set the seed to 0 for new images every run. On a large GPU, Images at once makes several images in one pass of the model; if the GPU runs out of memory, the block goes on one image at a time.
| Model | Size | Needs | Licence |
|---|---|---|---|
| Vega RT (Segmind) | 0.7B, 4 steps | Any worker; the fastest | Apache-2.0 |
| Vega (Segmind) | 0.7B, 25 steps | GPU with 8 GB | Apache-2.0 |
| SSD-1B (Segmind) | 1.3B, 25 steps | GPU with 8 GB | Apache-2.0 |
| Stable Diffusion XL 1.0 | 3.5B, 30 steps | GPU with 12 GB | CreativeML Open RAIL++-M |
| FLUX.1 schnell | 12B, 4 steps | GPU with about 40 GB and a token | Apache-2.0, gated |
auto picks Vega on a GPU or a Tenstorrent card and Vega RT on a CPU. Other takes any text-to-image model in the diffusers format, by its Hugging Face id (owner/name) or a folder on the worker; the block reads and records its licence.
Make videos
- Add Text to video (under AI) and a Video review view. Connect the block's live telemetry, its Videos and its Videos table to the view.
- Write one prompt per line: what happens, the camera, the light, the style. Set Videos per prompt: each variant starts from its own seed, so one prompt gives several different clips.
- Leave Seconds at 0 for the model's own length (5 s for most), or set up to 10 s. Clips longer than the model was trained for move less well.
- Press Run. Each video is sent whole as soon as it is made and plays in the view while the next one is being made.
| Model | Size and resolution | GPU memory | Licence |
|---|---|---|---|
| Wan 2.1 T2V (small) | 1.3B, 480p | 8 GB | Apache-2.0 |
| Kandinsky 5.0 Video Lite | 2B, 16 steps, 5 or 10 s | 12 GB | MIT |
| Wan 2.2 TI2V (default) | 5B, 720p | 24 GB (slowly on 8 GB, with 32 GB of computer memory) | Apache-2.0 |
| HunyuanVideo 1.5 | 8.3B, 480p | 14 GB or more | Tencent Hunyuan Community: not for the EU, the UK or South Korea |
| LTX-2.3 | 19B, 768 × 512, with sound, 30 steps | 80 GB, with 96 GB of computer memory | LTX-2 Community: free under US$10M of yearly revenue |
| Wan 2.2 T2V A14B | Two 14B experts, 720p | 80 GB, with 96 GB of computer memory | Apache-2.0 |
| LTX-2.5 (best) | 19B distilled, 960 × 544, with sound, 8 steps | 80 GB, with 96 GB of computer memory | LTX-2 Community, gated: needs a token |
LTX-2.3 and LTX-2.5 make the sound with the picture (steps, voices, wind, music) and put it in the MP4; the large player of Video review plays it. Your worker is a start; the largest models are made for cloud GPUs and for workstations with 48 to 96 GB cards.
A model larger than the GPU still runs: its layers wait in the computer's memory and move to the GPU group by group as they are needed. It is slower, and the log says so. File size: small keeps each MP4 under 4.5 MB, so it is uploaded to EsaSim and plays in the dashboard; high files stay on the worker and are fetched when you open them.
Which worker runs them
With Accelerator on auto or NVIDIA GPU, a Text to video run goes only to a worker with an NVIDIA GPU. If none is online, the run waits in the queue, and the Blueprint's console says so before you start. Choose CPU to let any worker take it. Text to image runs anywhere on auto; naming NVIDIA GPU, Tenstorrent or CPU sends it only to a worker that has one. A project can also limit its runs to some workers under Compute.
Review and keep the good ones
The Images view shows the pictures as a grid; click one to see it large and use the arrows to go through them. The Video review view shows each video as a card that plays under the pointer. Click a card to watch it large and whole.
| Key | In Video review |
|---|---|
| K / R | Keep or remove the video |
| U | Undo the decision |
| ← → | Previous or next video |
| Space / Esc | Pause, close the large player |
Filters show the videos still to decide, the kept ones or the removed ones. Decisions stay on your computer. Save the kept videos puts them, with their stills and a table of what made each, into a dataset (one new version) or into one zip. Save choices downloads choices.csv with the file, decision, seed and prompt of each video.
What each run saves
- The images (PNG or JPG) or videos (MP4, H.264) with a still of each.
images.csvorvideos.csv: seed, prompt, size, steps, time, model and licence of each file.summary.csv: one row per run (model, settings, device, seconds per file, loading time, licence). Join the summaries of several models in a Table or Chart view to compare them.generation.json: the settings, to make the same files again.
The output is not filtered: what it shows comes from your prompt. Check each model's licence before you use what it makes; some models restrict use or regions.
When something goes wrong
| Message or symptom | What to do |
|---|---|
| The run waits and no worker takes it | No worker with an NVIDIA GPU is online. Start one, or allow it under the project's Compute. |
| Needs an NVIDIA GPU, and this worker has none usable from PyTorch | The worker has no NVIDIA GPU, or its PyTorch was built without CUDA. Run on another worker, or set Accelerator to CPU for a test. |
| Not enough disk space | The block checks the free space before downloading. Free some space or pick a smaller model. |
| 401 or gated model | Accept the model's conditions on Hugging Face, then save your token in Profile > Keys. |
| The first run takes several minutes before the first file | The packages install and the model downloads once per worker; later runs start in seconds. |