# Pika H3 Flex

> Source: https://aiwiki.ai/wiki/pika_h3_flex
> Updated: 2026-11-01
> Fact-checked: 2026-11-01
> Categories: AI Models, Diffusion Models, Open Source AI, Video Generation
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Pika H3 Flex." aiwiki.ai, 1 Nov 2026. https://aiwiki.ai/wiki/pika_h3_flex
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Pika H3 Flex** is an open-weight video and audio generation model family released by [Pika](https://aiwiki.ai/wiki/pika_labs), the company operated by Mellis, Inc., on October 29, 2026. It is not a new architecture: Pika built it by post-training [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3), the open-weight video model from [MiniMax](https://aiwiki.ai/wiki/minimax), and it ships as two [LoRA](https://aiwiki.ai/wiki/lora) adapters applied on top of MiniMax H3's unchanged base weights.[1][2][3] Pika's post-training cut sampling from MiniMax H3's 50 denoising steps to six, which the company says brings transformer computation for a 15-second clip with audio below the clip's own running time (a measure that excludes encoding, decoding and serving overhead).[1][2] The open release covers text-to-video, first- and last-frame conditioning and multimodal reference-to-video at native 480p and 768p with mixed stereo audio; 1080p to 4K output, five separate audio stems and a prompt rewriter are hosted features of Pika's paid API.[1][3][6]

Pika announced the model as the top-ranked open-weight entry on [DesignArena](https://aiwiki.ai/wiki/designarena)'s text-to-video leaderboard, placing seventh overall at launch.[1][8] The adapters are distributed under the MiniMax H3 Community License Agreement, the same license as the base model, whose permitted territory excludes the European Union, the United Kingdom, South Korea and the United States.[3][15]

## Key facts

| Item | Detail |
| --- | --- |
| Developer | Pika (Mellis, Inc., Palo Alto, California)[3][23] |
| Released | October 29, 2026 (Hugging Face repository created 17:08 UTC; announced on X at 17:37 UTC)[3][8] |
| Base model | [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3) (Hugging Face MiniMaxAI/MiniMax-H3, pinned revision 42ed227)[3][17] |
| Release format | Two LoRA adapters (FL2VA and Ref2VA), FP32, rank and alpha 128; BF16 versions for [ComfyUI](https://aiwiki.ai/wiki/comfyui)[3] |
| Sampling | Six generator evaluations on a fixed re-noising schedule (teacher: Pika's fine-tuned 50-step MiniMax H3)[2][17] |
| Native output | 480p or 768p, 24 fps, 5 to 15 seconds, 11 aspect ratios, 32 kHz stereo audio[3] |
| Hosted extras | Upscaling to 1080p, 2K and 4K; five audio stems; prompt rewriter (Pika API only)[1][6] |
| References | Up to 12 files: 9 images, 3 videos, 3 audio clips[3] |
| License | MiniMax H3 Community License Agreement (adapters and base); Apache 2.0 (example code and the Qwen3-VL encoder)[3][16] |
| Launch API price | From $0.01 per output second (480p text- or image-to-video), as of November 1, 2026[6] |
| Distribution | [Hugging Face](https://aiwiki.ai/wiki/hugging_face), GitHub, ComfyUI, Hugging Face Diffusers, Pika API[1][12][13] |

## Background

Pika, co-founded by chief executive [Demi Guo](https://aiwiki.ai/wiki/demi_guo), is best known for its consumer video product, also called [Pika](https://aiwiki.ai/wiki/pika), and for in-house models such as [Pika 2.5](https://aiwiki.ai/wiki/pika_2_5). By 2026 the company also ran a developer marketplace, the Pika API Club, a $10-per-month membership that sells access to third-party video, image, audio and language models, MiniMax H3 among them, alongside Pika's own models.[7][24]

MiniMax H3 is distributed on Hugging Face under a community license dated August 2, 2026.[15] Its generator, the H3-Omni-Transformer, is described by MiniMax as a 33-billion-parameter dense single-stream transformer that predicts video and audio latents jointly, with the Qwen3-VL-32B vision-language model as its text and vision encoder.[14] MiniMax released the generator as two task-specific checkpoints, FL2VA (text and first/last-frame) and Ref2VA (multimodal references), and its model card says the H3-Regenerate-2K stage is not yet open-sourced, so local deployments of the open generator produce 768p output.[14] H3 Flex follows the same split: each Pika adapter targets one of those two checkpoints.

The relationship between the two models is documented in Pika's own release materials. The Hugging Face repository declares MiniMaxAI/MiniMax-H3 as its base model with an "adapter" relation, its NOTICE file describes the adapters as "Model Derivatives of MiniMax H3", and each adapter manifest records the pinned MiniMax H3 revision it was trained against.[3][16][17] The shared "H3" in the name refers to MiniMax's model; the third-party documentation from Comfy Org and Hugging Face Diffusers describes H3 Flex the same way, as a post-training of MiniMax H3 that introduces no new model architecture.[12][13]

## Release

Pika announced H3 Flex on October 29, 2026 with a post from its X account calling it "the #1 open weight model on DesignArena's t2v board" and promoting "exportable audio stems, resolutions up to 4k, and faster-than-realtime generation".[8] Chief executive Demi Guo described it the same day as "our open-weight video model built for creatives" with "more creative controls, leading quality, and cheapest cost".[9] The company published two blog posts that day, a launch post and a technical write-up credited to Pika Research, along with the Hugging Face weights repository (Pika-Labs/Pika-H3-Flex) and an inference repository on GitHub.[1][2][3][4]

Pika said at launch that the model was available through ComfyUI and Hugging Face Diffusers as well as its own API.[2] Comfy Org publishes three ComfyUI template workflows (text-to-video, reference-to-video and first-and-last-frame), and Hugging Face's Diffusers library supports the model as Modular Diffusers blocks on its main branch; the Diffusers-format copies of the MiniMax H3 base components in Pika's repository were contributed by Hugging Face.[12][13][16] Hosted access opened through the Pika API Club, and Pika said the model would come to its consumer creation product on November 2, 2026.[1]

Pika's launch post describes what is and is not open: "We're releasing Flex's generator adapters and inference code", covering text-to-video, first- and last-frame generation and multimodal reference-to-video with synchronized audio at 480p and 768p, while "the 1080p, 2K, and 4K upsampler, audio-stem separation, optional prompt enhancement, and accelerated serving stack are Pika-hosted features."[1] Pika's research post states that the company "will opensource the upsampler model", without giving a date.[2]

## Training

According to Pika's technical write-up, H3 Flex was produced in three post-training stages applied to the official guidance-distilled MiniMax H3 model.[2]

| Stage | Method | Purpose, as described by Pika |
| --- | --- | --- |
| 1. Supervised fine-tuning | Fine-tuning on a proprietary, in-house audio-video dataset "selected for aesthetic quality", spanning generation and editing tasks; the 50-step schedule was kept | Improve visual style and aesthetics; the result served as the teacher for distillation[2] |
| 2. Step distillation | Projected Distribution Matching Distillation (PDMD), after standard Distribution Matching Distillation proved unstable in Pika's experiments | Reduce joint audio-video sampling from 50 denoising steps to six[2][19][20] |
| 3. Post-distillation reward learning | Reward Feedback Learning (ReFL) and [Direct Preference Optimization](https://aiwiki.ai/wiki/direct_preference_optimization_dpo) (DPO), trained as two separate low-rank adapters and then merged into one | Improve aesthetics, clarity, instruction following and realism with no added inference steps[2][21][22] |

Distribution Matching Distillation (DMD) was introduced by Tianwei Yin and colleagues in 2023 as a way to distill a many-step diffusion model into a few-step or one-step generator by matching output distributions rather than individual sampling trajectories.[20] PDMD, by Zimo Wang and colleagues, is a variant aimed at video diffusion models, posted to arXiv in September 2026.[19] ReFL comes from the ImageReward paper (Xu et al., 2023), which fine-tunes a diffusion model directly against a learned reward model, and Diffusion-DPO (Wallace et al., 2023) adapts direct preference optimization to diffusion models.[21][22] Pika says human pairwise preferences were used to calibrate the automated judgments from which its DPO training pairs were built.[2]

Pika did not disclose the size or sources of its fine-tuning dataset; the write-up says only that it is proprietary, curated in-house and covers "everyday activities, performance, narrative, technical, and stylized content".[2]

## Architecture and release contents

H3 Flex reuses every component of MiniMax H3: the Qwen3-VL-32B conditioner, the video VAE, the audio VAE and the two generator partitions (transformer for FL2VA and transformer_ref for Ref2VA).[13][16] The base transformer configuration shipped in Pika's repository lists 50 transformer layers, a hidden size of 5,376, 56 attention heads of dimension 128 and a feed-forward dimension of 14,336.[18] Pika changes two things, in the words of the Diffusers documentation: "the sampler, a fixed six-step schedule that re-noises the clean prediction at every step in place of the Euler update, and the weights, a pair of rank-128 LoRA adapters over the two MiniMax-H3 transformer partitions."[13]

Each adapter manifest specifies the sampling contract: timestep knots at 1000, 750, 500, 250, 100, 40 and 0 (six intervals), a video shift of 12, an audio shift of 3, and a "renoise" method that draws fresh noise at intermediate steps. The adapters modify the attention query, key, value and output projections and the two feed-forward projections of each block.[17] Pika warns that "generic six-step Euler sampling is not equivalent" to this schedule.[3]

The Hugging Face repository contains:[3]

| Directory | Contents |
| --- | --- |
| fl2va/, ref2va/ | The two Pika adapters, FP32 LoRA, rank and alpha 128 (about 2.8 GB each) |
| comfy/loras/ | BF16 conversions of both adapters for ComfyUI's native MiniMax H3 layout |
| transformer/, transformer_ref/, text_encoder/, vae/, audio_vae/ | Diffusers-format MiniMax H3 base components, about 210 GB in total |
| examples/ | Apache 2.0 example scripts for Diffusers |

The repository's ComfyUI notes say the BF16 tensors were checked against the original adapters after conversion but that "an end-to-end ComfyUI render has not been verified as part of this upload".[3]

## Capabilities

The open-weight package supports five generation modes, all producing video with synchronized stereo audio:[3]

| Mode | Inputs | Adapter |
| --- | --- | --- |
| Text to video and audio | Prompt only | FL2VA |
| First-frame conditioning | Prompt and one opening frame | FL2VA |
| Last-frame conditioning | Prompt and one ending frame | FL2VA |
| First-and-last-frame conditioning | Prompt and both endpoint frames | FL2VA |
| Multimodal references | Prompt plus up to 9 images, 3 videos and 3 standalone audio clips (12 files total; at least one image or video) | Ref2VA |

Output is an MP4 file with H.264 video and AAC audio at 24 frames per second, in requested whole-second durations from 5 to 15 seconds. The model works on a 17n+5 frame grid, so a 5-second request yields 124 frames (about 5.17 seconds) and a 15-second request 362 frames (about 15.08 seconds). Supported aspect ratios are 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16 and 9:21, and prompts may run to 8,000 characters.[3] Reference videos and audio clips must each be 2 to 15 seconds long, with at most 15 seconds of combined video and 15 seconds of combined standalone audio.[3]

Pika's showcases emphasize prompts that script several camera moves or hard cuts with timestamps inside one 15-second clip, dialogue and sound design, motion graphics with on-screen lettering, product films built from a single product photo, motion transfer from a reference video onto a new character, and reference-guided editing of existing footage (swapping a setting, props or clothing while keeping a performance).[2][6] The model card cautions that these are "creative showcases, not release-validation renders or throughput benchmarks".[3]

### Hosted extensions

Three features are available only through the Pika API:[1][2][3]

- **Upscaling to 1080p, 2K and 4K.** Pika describes its upsampler as an in-house 1-billion-parameter transformer that works directly in pixel space and produces upscaled frames in a single network evaluation, trained from scratch on footage including native 4K video. Pika says the upsampler's first 1080p frame is available after 0.17 seconds on one GPU, with processing time scaling linearly with clip length.[2][2] In the API, a "draft to video" endpoint finishes a 480p or 768p draft at higher resolution and charges only the price difference.[6]
- **Audio stems.** A prompted source-separation model splits the generated soundtrack into five WAV stems: dialogue, score, SFX, foley and ambience. Pika calls this "a first for a video model"; the claim is the company's own.[2][6]
- **Prompt rewriter.** Included free with Pika-hosted generation and sold as a separate API for self-hosted users.[6]

## Inference and hardware

Pika's inference package, distributed as `pika-h3-pro`, runs one request on one GPU and supports pip, Conda and Docker installation with Python 3.12 and an NVIDIA GPU.[4][5] Its tested configurations all used a single [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100) 80GB; the documentation says H200 and B300 GPUs "have not been qualified". Default 480p, 5-second sampling allocated about 66 to 69 GiB of GPU memory in Pika's tests, while 768p, 15-second generation exceeded H100 memory and required CPU offload, which streams transformer blocks from host memory (about 29 to 32 GiB on the GPU and roughly 94 GiB of host memory observed).[5] A full download of both tasks takes about 210 GB of base components plus 5.5 GB of adapters.[4]

The separate Diffusers examples recommend at least 256 GB of host RAM per request when blocks are streamed from host memory.[3] Comfy Org's guide says a 5-second 1344x768 clip "finishes in about a minute on a single card" with the six-step schedule, compared with the 20 to 50 steps of a non-distilled H3 run.[12]

For its hosted service Pika reports further optimizations (fused kernels, cached timestep-dependent AdaLN projections, FlashAttention-4, block-sparse attention and compressed communication) that it says cut full inference for a five-second 768p clip with audio from 3.409 to 2.129 seconds on eight NVIDIA B300 GPUs.[2] Pika notes that local inference "will not necessarily match the optimized hosted API's speed".[6]

## Evaluation

### Pika's own benchmarks

All of the following results are self-reported by Pika and have not been independently reproduced. Pika compared H3 Flex at 768p with the official MiniMax H3, fal's hosted "H3 Max" and HeyGen Video (quality setting) on one prompt set, using seven automatic metrics.[2]

| Metric (higher is better) | Pika H3 Flex | MiniMax H3 | Fal H3 Max | HeyGen Video (quality) |
| --- | --- | --- | --- | --- |
| Video technical quality (DOVER++) | 12.3 | 11.7 | 10.6 | 11.5 |
| Video aesthetic (Aesthetic Predictor V2.5) | 5.41 | 5.28 | 5.00 | 5.33 |
| KVQ | 0.024 | 0.022 | 0.019 | 0.020 |
| UVQ | 0.602 | 0.595 | 0.590 | 0.595 |
| Audio aesthetic (Audiobox) | 7.02 | 6.85 | 6.89 | 6.76 |
| Audio-video alignment (ImageBind) | 0.212 | 0.208 | 0.216 | 0.201 |
| Text-video alignment (VideoCLIP-XL-v2) | 0.225 | 0.225 | 0.226 | 0.224 |

Source: Pika technical write-up, Table 3.[2]

H3 Flex scored highest on the five visual and audio quality metrics, while the alignment scores were close across all four models, and Fal H3 Max scored slightly higher on both alignment metrics.[2] In blind pairwise human preference studies against the same three models, Pika reports that H3 Flex led on human anatomy, complex composition and audio synchronization for text- and image-to-video, and on visual style and framing and camera motion for reference-to-video; Pika acknowledges that other systems were preferred for physics, multi-scene narrative, lighting and materials, and action-heavy or control-video reference tasks.[2] The write-up's opening claim that H3 Flex outperforms "leading competitors by up to 91%" matches its Figure 3, where H3 Flex scored better than Fal H3 Max on 91% of prompts for video aesthetic; the figure is a per-prompt win share on one metric against one model, not a 91% quality margin (its win shares on the other metrics ranged from 44% to 85%).[2]

On speed, Pika says transformer denoising time was below the clip duration in every configuration it evaluated, and that at 768p H3 Flex was about 1.8 times faster than Fal H3 Max on average across text-, image- and 10-second reference-to-video comparisons.[2][6] The headline claim that the model can generate "a 15 second clip in as little as 7 seconds" measures only transformer (DiT) computation for a 15-second 768p clip in a controlled benchmark; Pika states it is "not total end-to-end API latency".[1][6]

### DesignArena

[DesignArena](https://aiwiki.ai/wiki/designarena) ranks models with Bradley-Terry ratings from anonymous head-to-head user votes. Pika's launch post said that "at the time of writing" H3 Flex ranked "#1 among open-weights models and #7 overall" on DesignArena's text-to-video leaderboard, and its product page said it ranked above base H3, H3 Max and [Seedance 2.5](https://aiwiki.ai/wiki/seedance_2_5).[1][6] AlphaSignal reported an Elo rating of 1293 at the time of the announcement.[11]

The leaderboard moves as votes accumulate. As of DesignArena's update of November 1, 2026 (1:07 p.m. GMT+7), its video leaderboard showed:[10]

| Rank | Model | Organization | Elo | Win rate | Battles |
| --- | --- | --- | --- | --- | --- |
| 1 | Gemini Omni 1.1 Flash | Google | 1360 | 68.2% | 3,993 |
| 7 | Kling 4.0 | KlingAI | 1288 | 59.6% | 1,032 |
| 8 | Pika H3 Flex | Pika | 1288 | 59.5% | 817 |
| 9 | FLUX 3 Video | Black Forest Labs | 1285 | 64.5% | 7,571 |
| 10 | MiniMax H3 Max | Fal | 1277 | 53.5% | 20,804 |
| 11 | MiniMax H3 | MiniMax | 1276 | 63.7% | 5,851 |

H3 Flex's margin of error on that snapshot was plus or minus 3.4%, the widest among the top eight entries, reflecting its smaller number of battles.[10]

## Pricing

Pika charges per output-video second. Its product page lists a launch price at half of a stated regular price; as of November 1, 2026 the rates were:[6]

| Output | Text, image and keyframe to video (regular / launch) | Reference to video (regular / launch) |
| --- | --- | --- |
| 480p | $0.02 / $0.01 per second | $0.06 / $0.03 per second |
| 768p | $0.04 / $0.02 per second | $0.12 / $0.06 per second |
| 1080p | $0.07 / $0.035 per second | $0.15 / $0.075 per second |
| 2K | $0.08 / $0.04 per second | $0.16 / $0.08 per second |
| 4K | $0.16 / $0.08 per second | $0.20 / $0.10 per second |

When a video reference is used, input-video seconds are billed at the same rate as output seconds; image and audio references are free. Audio stems cost $0.001 per second and the standalone prompt rewriter $0.02 per request.[6][7] Pika offered API Club members 15 free H3 Flex generations per day from October 29 to November 5, 2026.[6] In its launch comparison table Pika listed H3 Flex's $0.01 starting price against $0.08 per second for MiniMax H3, $0.0325 for Fal H3 Max and $0.103 for Seedance 2.5, figures compiled by Pika.[2]

## License

The adapters are distributed under the MiniMax H3 Community License Agreement, dated August 2, 2026, as "Model Derivatives of MiniMax H3"; Pika's repository carries the required MiniMax notice alongside a copyright notice for Mellis, Inc. covering its own modifications.[3][16] The bundled MiniMax H3 base components keep the same agreement, while the Qwen3-VL encoder, tokenizer and processor assets are under Apache 2.0, as is Pika's example and inference code (which, Pika notes, does not cover the model weights).[3][4][16]

Terms of the MiniMax agreement that carry over to H3 Flex include:[15]

- The license applies only within an "Applicable Territory" defined as worldwide excluding the European Union, the United Kingdom, South Korea and the United States; use outside it is not authorized, and MiniMax invites parties in those regions to contact it about a separate license.
- Commercial products or services with more than $20 million in annual revenue need separate prior written authorization from MiniMax, and commercial products using the model must prominently display "MiniMax H3" in their user interface.
- The model and its outputs may not be used to improve any other AI model apart from MiniMax H3 and its derivatives.
- Anyone offering a hosted service built on the model must bind users to equivalent use restrictions and maintain safeguards and a reporting mechanism for violations.

Comfy Org's tutorial advises users to "review its terms, including the territorial and commercial-use conditions, before using outputs commercially".[12]

## Community adoption

Within three days of release, the Hugging Face repository had 275 likes and the GitHub inference repository 354 stars and 20 forks (as of November 1, 2026).[3][25] Third-party derivatives appeared on Hugging Face the same week, including a few-step "Turbo" LoRA, an Apple Silicon port using the MLX framework, and a quantized GGUF build for AMD ROCm hardware.[26] AlphaSignal's coverage described the release as an example of how "a team can specialize an existing video model, outperform its base on a preference leaderboard, and distribute the resulting adapters with inference code", while noting that local operation "carries a high storage and GPU cost".[11]

## See also

- [Pika](https://aiwiki.ai/wiki/pika)
- [Pika Labs](https://aiwiki.ai/wiki/pika_labs)
- [Demi Guo](https://aiwiki.ai/wiki/demi_guo)
- [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3)
- [DesignArena](https://aiwiki.ai/wiki/designarena)
- [Open weights](https://aiwiki.ai/wiki/open_weights)
- [Knowledge distillation](https://aiwiki.ai/wiki/knowledge_distillation)
- [Text-to-video generation](https://aiwiki.ai/wiki/text_to_video)
- [AI video generation](https://aiwiki.ai/wiki/ai_video_generation)

## References

1. Pika Team. "[Introducing Pika H3 Flex](https://pika.art/blog/pika-h3-flex)." Pika blog, October 29, 2026.
2. Pika Research. "[Pika H3 Flex Technical Deep Dive: Post-Training H3 to Six-NFE Omni-Generator](https://pika.art/blog/pika-h3-flex-research)." Pika blog, October 29, 2026.
3. Pika. "[Pika-Labs/Pika-H3-Flex](https://huggingface.co/Pika-Labs/Pika-H3-Flex)." Hugging Face model card and repository, created October 29, 2026. Accessed November 1, 2026.
4. Pika. "[Pika-Labs/Pika-H3-Flex: video and audio generation inference](https://github.com/Pika-Labs/Pika-H3-Flex)." GitHub repository, created October 29, 2026.
5. Pika. "[Hardware and tested configurations](https://github.com/Pika-Labs/Pika-H3-Flex/blob/main/docs/hardware.md)." Pika-H3-Flex GitHub documentation, October 2026.
6. Pika. "[Pika H3 Flex: Faster-than-real-time video model](https://pika.art/pika-h3-flex)." Product page. Accessed November 1, 2026.
7. Pika. "[Pika H3 Flex models](https://dev.pika.art/models/pika/pika-h3-flex)." Pika API documentation. Accessed November 1, 2026.
8. Pika (@pika_labs). "[Meet Pika H3 Flex](https://x.com/pika_labs/status/2115860471215788262)." X, October 29, 2026.
9. Demi Guo (@demi_guo_). "[Excited for Pika H3 Flex](https://x.com/demi_guo_/status/2115867589381632407)." X, October 29, 2026.
10. DesignArena. "[Video leaderboard](https://www.designarena.ai/leaderboard?tab=video&section=video&board=video)." Updated November 1, 2026. Accessed November 1, 2026.
11. AlphaSignal. "[Pika Drops H3 Flex, an Open Video Model Generating Audio 1.8x Faster](https://alphasignal.ai/news/pika-drops-h3-flex-an-open-video-model-generating-audio-1-8x-faster)." AlphaSignal, October 29, 2026.
12. Comfy Org. "[Pika H3 Flex ComfyUI Workflows](https://docs.comfy.org/tutorials/video/pika/pika-h3-flex)." ComfyUI documentation. Accessed November 1, 2026.
13. Hugging Face. "[Pika H3 Flex](https://huggingface.co/docs/diffusers/main/en/api/pipelines/pika_h3_flex)." Diffusers documentation (main branch). Accessed November 1, 2026.
14. MiniMax. "[MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3)." Hugging Face model card. Accessed November 1, 2026.
15. MiniMax. "[MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/42ed227ee7df40d41602854ae760620d6eb651fe/LICENSE)." August 2, 2026.
16. Pika. "[Components, sources and licenses (THIRD_PARTY_NOTICES.md)](https://huggingface.co/Pika-Labs/Pika-H3-Flex/blob/main/THIRD_PARTY_NOTICES.md)" and "[NOTICE](https://huggingface.co/Pika-Labs/Pika-H3-Flex/blob/main/NOTICE)." Pika-H3-Flex repository, October 2026.
17. Pika. "[FL2VA adapter manifest](https://huggingface.co/Pika-Labs/Pika-H3-Flex/blob/main/fl2va/manifest.json)." Pika-H3-Flex repository, October 2026.
18. Pika. "[Base transformer configuration](https://huggingface.co/Pika-Labs/Pika-H3-Flex/blob/main/transformer/config.json)." Pika-H3-Flex repository, October 2026.
19. Wang, Z., Yuan, J., Wang, A., Yang, H., et al. "[PDMD: Projected Distribution Matching Distillation for Video Diffusion Models](https://arxiv.org/abs/2609.35768)." arXiv:2609.35768, September 2026.
20. Yin, T., Gharbi, M., Zhang, R., Shechtman, E., et al. "[One-step Diffusion with Distribution Matching Distillation](https://arxiv.org/abs/2311.18828)." arXiv:2311.18828, 2023 (CVPR 2024).
21. Xu, J., Liu, X., Wu, Y., Tong, Y., et al. "[ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation](https://arxiv.org/abs/2304.05977)." arXiv:2304.05977, 2023 (NeurIPS 2023).
22. Wallace, B., Dang, M., Rafailov, R., Zhou, L., et al. "[Diffusion Model Alignment Using Direct Preference Optimization](https://arxiv.org/abs/2311.12908)." arXiv:2311.12908, 2023 (CVPR 2024).
23. Pika. "[Terms of Service](https://pika.art/terms-of-service)." Mellis, Inc. Accessed November 1, 2026.
24. Pika. "[Pika API Club pricing](https://dev.pika.art/pricing)." Pika API. Accessed November 1, 2026.
25. GitHub. "[Pika-Labs/Pika-H3-Flex repository metadata](https://api.github.com/repos/Pika-Labs/Pika-H3-Flex)." GitHub API. Accessed November 1, 2026.
26. Hugging Face. "[Models matching 'pika-h3-flex'](https://huggingface.co/models?search=pika-h3-flex)." Hugging Face Hub search. Accessed November 1, 2026.