In brief
- Pika H3 Flex’s weights became available for free on 30 October 2026, and the model can run in ComfyUI.
- In a test by one user on an RTX 3060 12 GB, total generation time for seven clips fell from 2,925 seconds to 2,415 seconds.
- In the test, the model preserved the reference image’s composition and background better than FastH3.
- Local generation is limited to 480p and 768p; upscaling to 1080p–4K is available through Pika’s paid API service.
Pika released the weights for Pika H3 Flex, a video-generation model based on MiniMax H3, for free on 30 October 2026. The model can be run locally through ComfyUI. A user test comparing Pika H3 Flex with FastH3 under the same conditions reports shorter generation times across seven clips and more consistent preservation of the scene layout in the reference images. The results are based on a trial by a single user on a computer with an RTX 3060 12 GB; each clip was generated only once.
What does Pika H3 Flex offer?
The model was not developed from scratch; Pika adapted MiniMax H3 with additional training. According to information reported in the source, videos with audio can be generated in six steps. It offers five generation methods: text-to-video, generation from the first or last frame, interpolation between two frames, and reference-based generation. Reference mode accepts up to 12 input items in total, with a limit of nine images, three videos, and three audio files.
Generated videos are 5–15 seconds long at 24 frames per second. Although the weights are free, local generation is limited to 480p and 768p resolutions. Upscaling to 1080p–4K, splitting audio into segments, and automatically rewriting prompts are available through Pika’s paid API service.
What did the seven-clip comparison show?
The user who published the test compared Pika H3 Flex with FastH3 using the same image, audio, prompt, 768 × 448 resolution, and seed value of 33. Pika H3 Flex generated all seven clips faster. Total generation time fell from 2,925 seconds to 2,415 seconds, a reduction of around 17% in the test.
Speed was not the only difference. In the comparison, Pika H3 Flex kept the house and surroundings in a snow-clearing scene, as well as the face of a character looking at a phone, closer to the reference image. FastH3 changed the background in some scenes or produced framing that diverged from the reference. However, motion and line-sharpness measurements favored FastH3 in six of the seven clips. The person who ran the test notes that these figures do not, on their own, describe the quality of the motion; for example, camera drift can also increase a motion measurement.
In the color comparison, Pika H3 Flex had lower saturation in all seven clips. But compared with the source images, Pika H3 Flex produced colors closer to the references in five of the seven clips. Since each clip was generated only once, the results should not be treated as a definitive measure of performance across different scenes or repeated runs.
Implications for architecture and visualization workflows
The tendency to preserve composition in the reference frame may be worth considering for visualization teams looking to turn a fixed interior or architectural image into a short video. This behavior could be useful, particularly when it is important to keep the scene’s background or framing unchanged. However, the source does not include a comparison using architectural scenes, so these results do not guarantee the same performance on architectural projects.
Local use requires ComfyUI and compatible model files. The person who ran the test reports completing generation on an RTX 3060 12 GB system that already had INT8-quantized H3 files and the text component installed, by adding two LoRA files. This single experience does not mean every computer will work the same way. The terms of the MiniMax H3 Community License should also be reviewed before starting a project, particularly regarding commercial use and permitted regions.
The ComfyUI setup notes cited in the source say that Pika H3 Flex should not be combined with acceleration methods such as those used by FastH3, and should be run at a fixed six steps. Rather than choosing a model for an entire office workflow in one go, a sounder starting point may be to test it on scenes where preserving the reference matters, while comparing alternatives for clips where motion and line sharpness are the priority.
Sources
2 sourcesSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
Pika H3 Flex’s practical value lies less in producing the most dynamic result in every scene and more in its tendency to preserve the composition and colors of the supplied image. That makes it an option worth testing for work that turns architectural visuals into short animations; however, the current comparison did not use architectural scenes and covers only seven one-off generations.
For offices in Turkey, free weights alone are not enough to make a decision. Local setup and hardware requirements should be tested in the workflow, and the current license terms should be reviewed carefully before commercial projects. The speed advantage is promising, but it is no substitute for scene-by-scene quality control.
Frequently asked questions
Is Pika H3 Flex free?
The model weights were released for free and can be used in ComfyUI. Some features, such as upscaling to 1080p–4K, are available through Pika’s paid API service.
Does Pika H3 Flex run on an RTX 3060?
The source reports that one user ran the model on an RTX 3060 12 GB system by adding two LoRAs to existing H3 files. This is not a general hardware guarantee for every setup.
Is Pika H3 Flex faster than FastH3?
In the reported seven-clip test, Pika H3 Flex generated every clip in less time, with total time down by around 17%. The comparison is based on a single user test.
