Learn the tool. Break the tool. Tell us everything.

Sign in to the beta site

Enter the email you used to request the video presets beta. We will send a one-time sign-in link, no password needed.

Your email Send sign-in link

If that address is on the beta list, a sign-in link is on its way. Open it on this device.
No email after a couple of minutes? That address may not be provisioned yet. Request access with the beta form or ping the WhatsApp group.

Continue in demo mode

The full conversation: Matt Gia and JT Visuals

The complete webinar walkthrough of every video model, plus the live demo of this beta. Timestamped chapters are on the YouTube page.

How to get access

1. Join the WhatsApp group. https://chat.whatsapp.com/ENk0itW3jYdCaXX2moi4i4
2. Request access with the beta form (also pinned in the group):
   Open the Video Presets Beta request form
3. Use your AutoHDR account email. No account yet? Start a trial at autohdr.com first.
4. We flip the beta on in your account, usually within a day.

From the webinar: the actual assets

Everything Matt and JT shared on the call. Small things download right here; big sets live in JT's Drive.

1. AI Video Prompting Workshop Guide (JT's full write-up, the basis of the presentation). Download the guide
2. Prompt Writers for every model (Matt's drop): ready-made prompt-builder skills for Veo, Kling, MiniMax H3, Seedance, Luma, Gemini Omni, plus a model-selector that picks the right model for your shot. Download the zip or get it from Drive
3. Model comparison footage (JT's Drive folder): the exact demo clips you see playing inside the webinar video above — the dark-sky shootout (all six generations plus per-model Kling / MiniMax / Seedance raws), the crowded-resort before-and-afters (Kling and Seedance, direct vs reference), and the original empty-plate footage every generation started from. Open the folder
4. The complete 62-slide presentation (everything Matt advanced through on screen, frozen as a self-contained HTML deck with every video playable): currently a 2.6 GB offline package — ask in the WhatsApp group and we will get it to you, or it gets a public link here the moment one exists.
5. JT's free community where he teaches all of this and more: skool.com/tsc-community
6. Magnific, the visualization tool shown during the demo: magnific.com

Using the tool inside AutoHDR

1. Open one of your listings. The left rail shows AutoHDR, Adjustments, Photo Models, and Video. Click Video. The modes are photo-to-video (animate a photo), Custom (the full video workspace), video-to-photo, and Extract Frame.
2. Pick a model on the left. Models are grouped by provider: Google, Kuaishou, ByteDance, MiniMax, Alibaba, Luma AI. Each card says what the model is best at and what frames it accepts.
3. Load your inputs. Drop in a start frame (a photo from the listing, or extract one from a video). Some models also take an optional end frame (Kling, Veo First-Last, H3) and reference images (Seedance 2.5 Reference, H3 Reference, Wan).
4. Write the prompt. Describe the motion and mood, not the whole scene. Granular on the creative works best: "sunbeams through the windows, slow forward push, natural shadows."
5. Set the parameters. Resolution (varies by model, 540p up to 4K), Aspect Ratio (16:9 for listing pages, 9:16 for reels), and Time. The Generate button always shows the credit cost for your current setup.
6. Read your transformation history on the right. Every generation keeps its source, output, model, and prompt in a lineage. See what worked, tweak one word, regenerate, compare.
7. Save the winner as a preset. That is the whole game: your proven model-plus-prompt-plus-settings chain becomes a reusable recipe. Coming soon to the beta: JT's own presets.

The JT and Matt method (from the webinar)

1. Run the same prompt on several models and compare. That is an unbiased test.
2. Pick the winner, then iterate the prompt on that model only.
3. Movement shots: MiniMax H3 Max or Seedance. Exact A-to-B transforms (day-to-night, snow-to-spring): Kling with start and end frames. Re-cropping real footage: Luma. Construction sequences: H3.
4. Iterate cheap, upscale last. Draft at low resolution, then run the winner at high res (Astra 2 or Topaz for final upscales).

Model reference

Verdicts are Matt's and JT's own words from the webinar. The complete research, with sources and timecodes, is the deep dive at the bottom of this page.

The complete research: every model, in depth

Three kinds of evidence: [UI] read from the AutoHDR beta interface, [WEBINAR] said on tape by Matt or JT (timecodes point into the full webinar video above), [WEB] public sources listed at the end.

Google: Veo 3.1 family and Gemini Omni Flash

In our UI: Veo 3.1, Veo 3.1 First-Last, Veo 3.1 Fast, Veo 3.1 Fast First-Last, Gemini Omni Flash Video, Gemini Omni Flash 1.1. Resolutions 720p / 1080p / 4K. [UI]

• [UI] Veo 3.1: "High-quality, cinematic videos with realistic motion." First-Last adds "precise first and last frame control." Fast is "at half the cost per second." Observed cost: 8 credits for a 4s 1080p generation.
• [WEB] Launched October 2025 by Google DeepMind: native audio, up to 4K output, first/last-frame control on both standard and Fast tiers. Public pricing runs roughly 0.35 to 0.40 dollars per second for standard and about 0.10 per second for Fast.
• [WEBINAR 00:12:59] Matt on Google's older video models: mostly replaceable by Kling. Omni Flash is the keeper: "gemini-omni-1.1-flash"-class, strong at editing existing footage and reference-to-video, 4K output, scenes extendable to about 40 seconds (the 1.1 entry in our UI adds the end frame and 4K tier).
• [WEBINAR 00:12:04] JT: "I honestly don't use Omni hardly at all. Most of my workflow is between H3, Seedance, and Kling." Use Omni for quick drafts and video edits, not hero shots.

Kuaishou: Kling 3.0 Pro and Turbo

In our UI: Kling 3.0 Pro ("Frontier Kling, with an optional end frame and clip length") and Kling 3.0 Turbo. 3 credits per generation. [UI]

• [WEBINAR 00:08:56] JT: "a trustworthy workhorse that I would use no matter what. It's had the best track record."
• [WEBINAR 00:11:09] Matt: cheap and fast with exact start/end adherence ("always starts on that start frame, always ends on that end frame"), but "the movements look a bit more fake, a little bit more AI."
• [WEBINAR 00:28:10] JT: movement can read clay-like ("Kling kind of feels like a Crayola..."), while other models put more realistic life into people in frame.
• [WEBINAR 01:01:08] Matt's rule of thumb: "Kling really shines when you give a start and end frame... quick, cheap transformations" (day-to-night, snow-to-spring). In the day-to-night shootout [01:00:41], Kling "didn't even fully get to nighttime" while H3 Max nailed it.
• [WEB] Kling 3.0 released February 2026: 3 to 15 second clips, first and last frame on both tiers, an All-in-One Reference mode (3-8s character video), claims of native 4K/60fps, and Turbo roughly 20x faster than standard.

MiniMax: the H3 family

In our UI: MiniMax H3 ("Image-to-video with an optional end frame"), H3 Max ("Sharper MiniMax H3, at 480P and 768P"), H3 Reference, and H3 Max Reference. [UI]

• [WEBINAR 00:55:47] Matt breaking down the family: H3 is the base, Max is the better-quality tier ("the most natural-looking is H3 Max" per JT [00:57:10], whose base H3 can have a stop-motion feel).
• [WEBINAR 01:00:25] The day-to-night shootout went to H3 Max; Matt: "the movement is so much more natural" (Kling "is not trained on, clearly, like, actual enough real video").
• [WEBINAR 01:02:49] JT on construction and build sequences: "H3 is obviously the most construction-like" (the drywall and tape-lines detail held up).
• [WEBINAR 00:09:27] Matt on cost: top models ran him about 20 dollars for high-res generations; H3 is the cheaper, faster tier.
• [WEB] H3 is natively multimodal with synced stereo audio, 5-15 second clips, 768p and 2K tiers, keyframe and reference inputs. Our Reference variants add reference-image conditioning.

ByteDance: Seedance 2.5

In our UI: Seedance 2.5 ("Next-generation image-to-video, up to 30 seconds") and the Reference variant for longer, reference-driven clips. [UI]

• [WEBINAR 00:12:04] Part of JT's daily trio (H3, Seedance, Kling).
• [WEBINAR 01:01:08] Matt: use H3 and Seedance for movement shots, Kling for the exact A-to-B transforms.
• [WEB] Released July 2026: one-take 30-second audio-video clips, roughly 30 to 50 multimodal references (images, video, audio), 1080p to 4K output, region editing, identity/portrait reference support.

Alibaba: Wan 3.0 and Wan 3.0 Prime

In our UI: Wan 3.0 ("Clips up to 30 seconds, with an optional end frame") and Wan 3.0 Prime ("tuned for identity and scene continuity"). [UI]

• [WEB] Announced August 2026: 2 to 30 second single-pass generation with native audio and multimodal reference inputs (turning even text-heavy material into video); Prime is the premium tier tuned for identity and scene continuity, strongest in reference-to-video workflows.
• No on-tape verdict from Matt or JT in our webinars yet. Try it where a shot needs to hold a face or a room's identity across a longer clip, and report back through the feedback form.

Luma AI: Ray 3.2

In our UI: Luma Ray 3.2 ("Cinematic five-second motion from a start frame"). 540p / 720p / 1080p, every aspect ratio. 5 credits. [UI]

• [WEBINAR 00:24:29] JT: "I like the movement from Luma, but it's a little shaky... Kling 3 is nice and cinematic and stable."
• [WEBINAR 00:11:28] Matt: "for advanced users only," but it can reference exact keyframes along the timeline, which no other model here does.
• [WEBINAR 00:37:11] Matt's production use: re-cropping real horizontal footage into vertical without hallucinating new pixels.
• [WEB] 5 or 10 second clips, aspect ratios from 9:16 up, start and end frame interpolation, and a Reframe feature for changing aspect after generation.

Adjacent tools from the webinar (not in our picker)

• Astra (GPT with vision) [WEBINAR 00:31:39]: JT's orchestration assistant; it reads footage frame by frame and writes the prompts. His biggest cheat code [01:39:23]: "have AI read the documentation... You just focus on being creative."
• Astra 2 upscaler [WEBINAR 01:39:01]: Matt calls it "the best upscaler for video I've found right now."
• Topaz [WEBINAR 01:38:32]: JT: longest time in market, enterprise-grade, movement-preserving upscaling.
• Nano Banana (Google image model) [WEBINAR 01:43:01]: for wild re-angles on the photo side; AutoHDR's own models are fine-tuned on real estate to not hallucinate.
• LTX 2.5 [WEBINAR 00:07:24]: named as a recent major release; not in our picker yet.

Sources

Google: Veo 3.1 announcement, Gemini API pricing, DeepMind. Kling: kling.ai, Pexo on 3.0 Turbo. MiniMax: minimax.io, Hailuo. Seedance: ByteDance Seed announcement, fal.ai. Wan: Alibaba Cloud, Runway. Luma: lumalabs.ai, Scenario. Webinar quotes: the full JT and Matt conversation embedded at the top of this page.