Pika · research post 2026-10-29 · read 2026-11-01
Pika H3 Flex: six-step LoRAs on MiniMax H3
Pika H3 Flex is Pika's post-training of MiniMax H3 into two six-step LoRA adapters, FL2VA and Ref2VA, that run on the H3 base in ComfyUI.
This page is not Pika. It does not load Pika H3 Flex LoRAs, does not call the Pika API, and the clip is Hailuo 02 Standard for 60 credits, 6 seconds.
Duration
Ratio
Generated Videos
No videos yet
Fill out the form and click Generate Video to create your first AI video.
Tips: Use detailed descriptions for better results
- Made by
- Pika
- Released
- October 29, 2026
- Run on Nano AI
- Hailuo 02 Standard · from 60 credits per video
- Page updated
- November 1, 2026
Source: pika.art
What Pika H3 Flex is
Pika H3 Flex is a post-trained version of MiniMax H3. The ComfyUI tutorial, read on 2026-11-01, says Pika fine-tuned H3 on a curated multi-task corpus of video, audio and reference examples, distilled the 50-step generator down to six steps, then refined it with reward feedback and preference optimization. It ships as two LoRA adapters on the MiniMax H3 base, loaded through LoraLoaderModelOnly. The tutorial says no new model architecture is involved. Pika's research post is dated 2026-10-29.
The Hugging Face card, also read on 2026-11-01, separates local open weights from the hosted API. Local inference is native 480p / 768p with mixed 32 kHz stereo, 24 fps, and requested lengths of 5-15 seconds. A 5-second request is 124 frames, about 5.17 seconds, and a 15-second request is 362 frames, about 15.08 seconds. The card lists 11 aspect ratios: 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16 and 9:21. Output is an MP4 with H.264 video and AAC audio, plus JSON metadata. Prompts are 1 to 8000 characters. The hosted API adds 1080p, 2K and 4K draft-to-video, audio stems and prompt enhancement. Those are not in the downloadable adapters. The card header lists 33B parameters. Base components are about 210 GB in total, and the adapters are FP32 LoRA rank/alpha 128. The Qwen encoder and the example code are Apache 2.0; that code license does not cover the weights.
Nano AI does not run those adapters and does not call the Pika API. The form on this page runs Hailuo 02 Standard, a different MiniMax clip model this site has priced: 6 seconds, 60 credits. Eligible new accounts receive 64 credits, which covers one such clip. The Mini Pack is $6.99 for 400 credits. For the open H3 base, including what is left out of the weights, see the MiniMax H3 page on this site.
Pika H3 Flex at a glance
- Vendor
- Pika, post-trained on MiniMax H3
- Research post
- 2026-10-29
- Pages read
- 2026-11-01
- Adapters
- FL2VA and Ref2VA LoRAs
- Local picture
- Native 480p / 768p; 11 ratios: 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3, 9:16, 9:21
- Local audio
- 32 kHz mixed stereo, H.264 and AAC in the MP4
- Sampling
- 6 steps: lcm, ManualSigmas, CFG 1.0, strength 1.0
- ComfyUI
- T2V 0.43.2+; Ref2VA and FL2VA 0.30.0+
- On Nano AI
- Hailuo 02 Standard, 6 seconds, 60 credits
Where the ComfyUI files go
File names and folders are from the ComfyUI Pika H3 Flex tutorial as read on 2026-11-01. The H3 base files are shared with the MiniMax H3 templates. The only new files are the two adapters. The tutorial does not list a per-file byte size, so this page does not invent one. Nothing here is downloaded onto Nano AI.
- 1
Match the ComfyUI version to the template
Text to Video needs ComfyUI 0.43.2 or newer. Reference to Video and First and Last Frame need 0.30.0 or newer. All three are under Template Library > Video. The tutorial calls the family video_pika_h3_flex_*.
- 2
Place the shared base and the two adapters
Diffusion models, the two LoRAs, the text encoder and the two VAEs use the folders below. Search Pika H3 Flex: Text to Video, Pika H3 Flex: Reference to Video, or Pika H3 Flex: FL2VA if you would rather load a template than place files by hand.
ComfyUI/models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors ComfyUI/models/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors ComfyUI/models/loras/pika_h3_flex_fl2va_v1_bf16.safetensors ComfyUI/models/loras/pika_h3_flex_ref2va_v1_bf16.safetensors ComfyUI/models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors ComfyUI/models/vae/minimax_h3_video_vae_int8_convrot.safetensors ComfyUI/models/vae/minimax_h3_audio_vae_fp32.safetensors - 3
Keep the six-step schedule
Sampler lcm, scheduler ManualSigmas, CFG 1.0 through BasicGuider, adapter strength 1.0. Do not stack a turbo or Lightning LoRA. Do not drop the sigma list below 7 values.
- 4
Read the license before commercial use
The adapters are under the MiniMax H3 Community License, including territorial and commercial-use conditions. The card also names the Mellis, Inc. copyright statement in NOTICE.
Files the three templates name
Names from the ComfyUI tutorial layout, read on 2026-11-01. Base rows link to the Comfy-Org repack, which lists those diffusion, text-encoder and VAE file names. Adapter rows link to Pika-Labs/Pika-H3-Flex, where the card says the Comfy-format BF16 LoRAs live under comfy/loras/. The tutorial also says Comfy republishes them in Comfy-Org/MiniMax-H3 under the same names.
| File | Folder | Template | Role |
|---|---|---|---|
| minimax_h3_fl2va_pruned_int8_convrot.safetensors | ComfyUI/models/diffusion_models/ | T2V and FL2VA | H3 diffusion checkpoint the templates name |
| minimax_h3_ref2va_pruned_int8_convrot.safetensors | ComfyUI/models/diffusion_models/ | Ref2VA | H3 diffusion checkpoint the templates name |
| pika_h3_flex_fl2va_v1_bf16.safetensors | ComfyUI/models/loras/ | T2V and FL2VA | Flex adapter, BF16 file name from the tutorial |
| pika_h3_flex_ref2va_v1_bf16.safetensors | ComfyUI/models/loras/ | Ref2VA | Flex adapter, BF16 file name from the tutorial |
| qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | ComfyUI/models/text_encoders/ | All three | Text encoder shared with the H3 templates |
| minimax_h3_video_vae_int8_convrot.safetensors | ComfyUI/models/vae/ | All three | Video VAE |
| minimax_h3_audio_vae_fp32.safetensors | ComfyUI/models/vae/ | All three | Audio VAE |
Before you run the adapters
The license is MiniMax's
The ComfyUI tutorial and the Hugging Face card say the adapters are distributed under the MiniMax H3 Community License, including territorial and commercial-use conditions. The card calls them Model Derivatives of MiniMax H3 and points at NOTICE for the MiniMax notice and the Mellis, Inc. copyright statement. Read that text before you use outputs commercially. This website does not grant that license.
Do not stack a turbo LoRA
The tutorial ships Pika's six-step schedule: lcm with ManualSigmas and CFG 1.0 through BasicGuider. Keep the sigma list and the adapter strength at 1.0. Do not stack a turbo or Lightning LoRA on top. The adapter is already distilled to six steps, and a second acceleration LoRA breaks the output. Lowering the sigma list below 7 values breaks the schedule, and raising the step count does not add detail. The Comfy-Org/MiniMax-H3 card, read the same day, lists separate turbo files in loras/, including minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors, minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors and minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors. Those are not the Flex adapters.
4K and stems are API extensions
Local open adapters are native 480p / 768p with mixed stereo. The Hugging Face card says 1080p, 2K and 4K draft-to-video, audio stems and prompt enhancement are hosted endpoints, separate from the downloadable adapters. The 2026-10-29 post names five stems: dialogue, score, SFX, foley and ambience. That post's 4K mark describes the product family, not a file in the local layout.
The base is the open MiniMax H3
ComfyUI says the H3 base files are shared with the existing MiniMax H3 templates, and the only new files are the two adapters. Those templates use res_multistep with BasicScheduler, the 20-step path. The MiniMax H3 page on this site records the open-base limits already published there: H3-Context-IR and H3-Regenerate-2K are not in the open weights. This page does not retest that base.
Local adapters, the Pika API, MiniMax H3, and Fal's listed prices
Local and API cells quote the Hugging Face card and the ComfyUI tutorial as read on 2026-11-01. Fal H3 Max and Fal H3 Turbo appear only as Pika printed them in Table 1 and Table 2 of the research post dated 2026-10-29. Table 1 prices are USD per output-second. Table 2's header says USD per second; the cells below are copied as printed, without adding a dollar sign. This site did not measure Fal, and it does not sell those rates.
| Local Comfy / open adapters | Pika hosted API | MiniMax H3 | Fal H3 Max | Fal H3 Turbo | |
|---|---|---|---|---|---|
| Picture | Native 480p and 768p | 1080p, 2K and 4K by draft-to-video | 768p in the open base; 2K only through closed H3-Regenerate-2K | Table 1: 4K marked no | Table 1: 4K marked no |
| Sampling | 6 steps: lcm, ManualSigmas, CFG 1.0, adapter strength 1.0 | 6-step family; the faster stack is served through the API | Comfy H3 templates use a 20-step res_multistep path; Pika distilled a 50-step teacher | Not stated in Table 1 or Table 2 | Not stated in Table 1 or Table 2 |
| Audio | 32 kHz mixed stereo in the MP4 | Stems endpoint; the post names dialogue, score, SFX, foley and ambience | 32 kHz stereo on the open card; Table 1 marks stems no | Table 1: stems marked no | Table 1: stems marked no |
| Text or image price | No hosted rate in these sources | Table 1: $0.01 per output-second | Table 1: $0.08 per output-second | Table 1: $0.0325 per output-second | Table 1: $0.01625 per output-second |
| Reference price | No hosted rate in these sources | Table 2: 0.03 per second | Table 2: 0.08 per second | Table 2: 0.05 per second | No Table 2 row; the post says no reference-to-video endpoint |
| Open weights | Two LoRAs: FL2VA and Ref2VA | Draft-to-video, stems and prompt enhancement are separate from the adapters | Open FL2VA and Ref2VA base; Table 1 marks open weights yes | Table 1: open weights marked no | Table 1: open weights marked no |
| Wait, as quoted | Tutorial: about one minute for a 5-second 1344x768 clip on one card | Table 1 marks real-time yes | Table 1 marks real-time no | Table 1 marks real-time yes | Table 1 marks real-time yes |
| On Nano AI | Not loaded | Not called | Not run here | Not run here | Not run here; the button is Hailuo 02 Standard for 60 credits |
Primary sources
Read these for anything this page does not quote. All three were read on 2026-11-01. The research post is dated 2026-10-29.
Shortest ComfyUI path: three templates
- 1
Text to Video
Prompt only, no image input. In Template Library > Video, search Pika H3 Flex: Text to Video. ComfyUI 0.43.2 or newer. The docs name the workflow family video_pika_h3_flex_*. It loads minimax_h3_fl2va_pruned_int8_convrot and the LoRA pika_h3_flex_fl2va_v1_bf16, plus the shared text encoder and both VAEs. The Hugging Face card calls the matching CLI task t2va. First-frame-only and last-frame-only are the card's i2va and l2va tasks; this template is the prompt-only one.
- 2
Reference to Video
Two or more image references carry identity, props and scene. The prompt carries the action. Search Pika H3 Flex: Reference to Video. ComfyUI 0.30.0 or newer. It loads minimax_h3_ref2va_pruned_int8_convrot and pika_h3_flex_ref2va_v1_bf16. The card's CLI task is ref2va. Limits, from the tutorial and the card: up to 9 images, 3 videos and 3 audio clips, 12 files in total. Each video or standalone audio clip is 2-15 seconds, and the combined duration of each type stays within 15 seconds. At least one image or video is required. A video's embedded soundtrack does not count as another audio file. The card says these are input limits, not a promise that every maximum fits in GPU memory.
- 3
First and Last Frame
Two frames anchor the opening and the ending. Search Pika H3 Flex: FL2VA. ComfyUI 0.30.0 or newer. It uses the same diffusion file and the same LoRA as text-to-video: minimax_h3_fl2va_pruned_int8_convrot and pika_h3_flex_fl2va_v1_bf16. The card's CLI task for both endpoint frames is fl2va. All three templates keep the six-step schedule. Do not add a turbo or Lightning LoRA.
Prompt habits, then a 6-second Hailuo clip
Paraphrased from the ComfyUI prompting notes read on 2026-11-01, not copied from the long showcase prompts. For T2V and FL2VA, describe the style, then each shot with its camera move, then the two sound fields. For Ref2VA, the full-reference fields named there are subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape and non_diegetic_music. Camera amplitude is advisory: put the move in the shot sentence and give it a stop. Explicit degrees and cut timestamps can land 0.2 to 0.3 seconds off, or the other way. Wide shot alone tends to become a medium shot; say the subject is small in a large empty room. A hand meant to stop a few centimetres away usually makes contact. Short labels on props tend to stay legible. The four lines below are for Hailuo 02 Standard on this page. They are not Pika H3 Flex prompts.
Wide empty basin
A small red paper boat sits in the middle of a large empty concrete basin, with a wide stretch of bare floor around it. The camera moves slowly closer and stops at the basin rim. Soft daylight, no people, no readable text, no logos.
Lantern, then hold
A paper lantern hangs in a still stone courtyard. The camera arcs partway around the lantern and then holds still. Dusk, no people, no readable text, no logos.
Ball against a cup
A wooden ball rolls across a table and comes to rest against a ceramic cup. Side light, the camera holds still, no hands, no readable text, no logos.
Short bottle label
A matte black bottle on a white table, with a short label that reads OAK. The camera pushes in slowly and stops. No people, no other text, no logos.
If the run is not the product you expected
- A turbo LoRA was stacked on the adapter
- Take it off. The tutorial says the Flex adapter is already distilled to six steps, and a turbo or Lightning LoRA breaks the output. Stay on lcm, ManualSigmas, CFG 1.0 and strength 1.0.
- You expected 4K or separate stems from the local files
- Those are Pika API extensions: draft-to-video for 1080p, 2K and 4K, plus an audio-stems endpoint. The local adapters are 480p / 768p with mixed stereo.
- The H3 template is still on the 20-step scheduler
- The ComfyUI FAQ says to load the Pika adapter, switch the sampler to lcm, and replace BasicScheduler with ManualSigmas. res_multistep plus BasicScheduler is the 20-step path, not the Flex schedule.
- This website still says Hailuo
- That is the priced model on the button: Hailuo 02 Standard, 6 seconds, 60 credits. It does not load Pika H3 Flex LoRAs and does not call the Pika API.
FAQ
Pika H3 Flex FAQ
What the ComfyUI tutorial, the Hugging Face card and the 2026-10-29 post say, and what the button on this page calls.
Does Nano AI run Pika H3 Flex?
No. This page is not Pika. It does not load Pika H3 Flex LoRAs and does not call the Pika API. The generator runs Hailuo 02 Standard for 60 credits and 6 seconds. Eligible new accounts receive 64 credits, which covers one clip. The Mini Pack is $6.99 for 400 credits.
Where are the weights?
Pika releases them at Pika-Labs/Pika-H3-Flex. The ComfyUI tutorial says Comfy republishes the ComfyUI-ready files in Comfy-Org/MiniMax-H3 under the same file names, and the template nodes point at that repack. The base components are MiniMaxAI/MiniMax-H3. The card lists FP32 LoRA rank/alpha 128 under fl2va/ and ref2va/, and Comfy-format BF16 LoRAs under comfy/loras/. This website does not host those files.
Why six steps?
The tutorial and the 2026-10-29 post say Pika distilled the 50-step H3 generator with distribution matching distillation, then refined it with reward feedback. Six steps is the trained schedule. Raising the step count does not add detail. Lowering the sigma list below 7 values breaks the schedule. Keep lcm, ManualSigmas, CFG 1.0 through BasicGuider, and adapter strength 1.0. Do not stack a turbo or Lightning LoRA.
Can I use the adapters on MiniMax H3 templates?
Yes, according to the ComfyUI FAQ read on 2026-11-01. The MiniMax H3 templates use res_multistep with BasicScheduler, which is the 20-step path. Load the Pika adapter in the template's LoRA node, switch the sampler to lcm, and replace the scheduler with ManualSigmas. The node the tutorial names is LoraLoaderModelOnly. The open base still omits H3-Context-IR and H3-Regenerate-2K, as the MiniMax H3 page on this site already says.
Do the local adapters output 4K or separate stems?
No. The Hugging Face card says local open-weight inference is native 480p / 768p with mixed stereo audio. 1080p, 2K and 4K draft-to-video, audio stems and prompt enhancement are Pika API extensions. The 2026-10-29 post names the five stems dialogue, score, SFX, foley and ambience, and Table 1 marks 4K for Pika H3 Flex. Treat that 4K mark as the hosted product, not as a file in the ComfyUI layout.
Is Nano AI affiliated with Pika?
No. Nano AI is not affiliated with Pika, MiniMax, Hailuo or Fal. The clip from the button is Hailuo 02 Standard.
More from Nano AI
Related pages
Nano AI is an independent service and is not affiliated with Pika, MiniMax, Hailuo or Fal. Pika H3 Flex facts are from the ComfyUI tutorial, the Hugging Face model card, and Pika's research post dated 2026-10-29, as read on 2026-11-01. The generator runs Hailuo 02 Standard.