Creative AI Briefing October 30, 2026 morning edition 11 min read

The brief is becoming the input, so the time a creator spends gathering, labelling and checking source material now decides the result more than the wording of a prompt.

Pika H3 Flex MiniMax H3 ElevenLabs SFX v3 Krea Ad Studio SnapRef video-gen audio-gen open-weights comfyui licensing-provenance design-tools

Creative AI Briefing: Friday, October 30, 2026

A filmmaker can now hand a video model up to nine photos, three clips and three sound files, name each one in the prompt, and get back a 5 to 15 second shot with synced sound in six passes instead of fifty. That is Pika H3 Flex, an open-weights model Pika published yesterday. ElevenLabs shipped a sound effects model built to follow several sound instructions in one prompt, and Krea shipped a tool that reads a shop's whole website before it draws a single ad. The pattern: the brief is becoming the input. The hour you spend collecting, labelling and checking your material now decides the output more than any clever sentence does, and every one of these tools says so in its own documentation.

New models

Pika H3 Flex: Pika's open video-and-sound model, in six steps. Pika-Labs/Pika-H3-Flex (created October 29, 17:08 UTC, per the Hugging Face API; launch post, technical post) is a retrained version of MiniMax H3, the open model that makes picture and sound together, cut down to six generation passes from the 50 its teacher uses. It ships as two add-ons: one for text plus optional first and last frames, and one for references, where you can supply up to 12 files (9 images, 3 videos, 3 audio clips, each clip 2 to 15 seconds). At least one image or video is required, and combined video and combined audio are each capped at 15 seconds. Output, per the model card, is 480p or 768p in 11 aspect ratios, 5 to 15 seconds at 24 fps with 32 kHz stereo sound. ComfyUI has official templates (video_pika_h3_flex_t2v.json, video_pika_h3_flex_fl2va.json, video_pika_h3_flex_ref2va.json) and the inference code sits on GitHub (150 stars, shields.io). The catches are real. Pika's own hardware notes record roughly 66 to 69 GiB of graphics memory for a 480p, 5-second clip on an H100, and 768p at 15 seconds only fit with CPU offload, at about 29 to 32 GiB on the card with host memory observed near 94 GiB including page cache (Pika says these are not minimum requirements); the base components want at least 211 GB of disk. Comfy's tutorial says a 5-second 1344 by 768 clip finishes "in about a minute on a single card" without naming the card. The quality comparisons in the technical post are Pika's own, and by its own account rivals were preferred on physics, multi-scene stories, lighting and action-heavy reference shots. Then the licence: the add-ons are MiniMax H3 derivatives under the MiniMax H3 Community License, which excludes the United States, the European Union, the United Kingdom and South Korea, forbids using the works "or any of their Outputs" there, and requires MiniMax's written permission above US$20 million in yearly revenue. Pika is a US company; its home market cannot run these weights under that licence. The hosted route is the Pika API, from $0.01 per output second for text or image to video and $0.03 per second for reference to video; the API also offers 1080p, 2K and 4K finishing and five separate sound stems (dialogue, score, effects, foley, ambience) as hosted features, with no separate prices published. A Pika-Labs/pika-h3-flex Space is running on Hugging Face's shared ZeroGPU hardware.

ElevenLabs SFX v3: a sound scene from one prompt. SFX v3 (October 29) is ElevenLabs' third sound effects model, and the headline change is that it follows "several instructions within a single prompt": a door, the rain behind it and the dog that barks at it, in one generation. ElevenLabs also says gongs, chimes and whistles come out with fewer artifacts, and its docs add that music and vocal sounds are handled better. It runs inside ElevenCreative (Sound Effects tab, Flows, Studio, Music Editor), the mobile app and the API as eleven_text_to_sound_v3. The docs give the working limits: 0.1 to 30 seconds per generation, MP3 or 48 kHz WAV for non-looping effects, a loop option, and a "prompt influence" control from literal to loose. Setting the duration yourself costs 40 credits per second. The catches, from the docs: although the blog says "available now," the API model is in limited access and "may not be available in every region"; exact counts like "triangle hit 4 times" are not always followed; small wording changes swing results; the docs describe looping in terms that conflict with the 30-second cap. The benchmark improvement is ElevenLabs' own. Try it at elevenlabs.io/sound-effects.

Image

Krea Ad Studio: a brand kit first, then the ads. Ad Studio (October 29) reads a website, builds an editable brand kit (logo, colours, fonts, tone, audience, selling points, up to 20 products) and generates static ads in batches of up to 12, sized for Instagram feed (4:5, 1:1), Stories and Reels (9:16) and landscape feed (16:9). Krea calls it "the first Krea 3 Expert." On one timed run on a real store, Krea reports a brand kit in 49 seconds and six Story ads in 2 minutes 28 seconds. Shopify catalogues import directly. It is on every paid Krea plan from Basic; brand research and product import are free, ads spend compute units. The catch is in Krea's own instructions: the ad copy comes from the selling points in the kit, so a sloppy kit becomes twelve sloppy ads. Krea describes the output as static ads.

SnapRef: one character picture, new poses, non-commercial. 0ppxnhximxr/snapref-v1b (created October 29) is a reference add-on for the Anima base model: give it one image of a character and your prompt sets pose and scene, with a mode that copies the character only or the character plus its drawing style. It ships its own ComfyUI node and example workflows, and the author reports about 9.9 seconds per 832 by 1216 image at 28 steps on an A100. Licence is the CircleStone Labs Non-Commercial License, so this is for personal work and pitches. The card itself warns that earrings and scars get simplified, dark-skinned references can come out lighter, and two characters swap attributes. A smaller, slightly faster sibling, snapref-v1a, landed the same day, and its card suggests starting there.

Video

The Pika model, cut further by the community within hours. WAS/Pika-H3-Flex-PTD (created October 29, 18:21 UTC) merges Pika's add-on into MiniMax H3's four-step files at half strength, as two roughly 4.8 GB files for WAS Node Suite's H3 PTD Patch node; no memory figures. akhaliq/Pika-H3-Flex-Turbo-LoRA (created October 29, 18:45 UTC) is tagged for 2, 4 and 8 steps. Both inherit the same H3 licence. Per ComfyUI Wiki's launch roundup, an early tester relayed that WAS saw fine detail "popping" at four steps and recommended staying at six, and Pika's adapter notes say "no end-to-end ComfyUI render was verified as part of the upload," so expect some drift from Pika's own results.

Audio and music

Beyond SFX v3, the agentory account posted GGUF repacks of YuE2, MiniMax Music 3, Stable Audio 3 Small Music and MOSS SoundEffect on October 28 and 29: second-day packaging, no new capability.

Open and local

The local story is a hardware wall. Pika shipped ComfyUI templates on day one, but its own tested configurations start at a data-centre card. The community had faster variants out within hours, the usual first step toward smaller cards; we found no consumer-GPU run published yet.

Pika-Labs/Pika-H3-Flex: six-step video with sound, up to 12 references; created October 29, MiniMax H3 licence with territory exclusions (card).
Pika-Labs/Pika-H3-Flex (GitHub): the local CLI, prompt guide and hardware notes; 150 stars, shields.io (repo).
WAS/Pika-H3-Flex-PTD: Pika's add-on merged into four-step H3 files for ComfyUI; created October 29 (card).
akhaliq/Pika-H3-Flex-Turbo-LoRA: fewer-step variants for fast drafts; created October 29 (card).
0ppxnhximxr/snapref-v1b: one picture of a character, new poses, ComfyUI node included, non-commercial; created October 29 (card).
lvladikov/Krea2-Turbo-Distill-3step-LoRA: a three-step Krea 2 Turbo add-on, marked by its author as still in training; created October 28 (card).
elayadesign/ai-design-skills: a landing-page design skill file for Claude Code, Cursor and Codex, MIT; on Trendshift's daily list this morning (repo).

Creative workflows

1. Plan a shot between two frames with Pika H3 Flex. Source: Comfy's Pika H3 Flex tutorial, template video_pika_h3_flex_fl2va.json in Comfy-Org/workflow_templates, plus Pika's prompt guidance.

The steps. Open Template Library, Video, First and Last Frame. Place minimax_h3_fl2va_pruned_int8_convrot.safetensors in models/diffusion_models/, pika_h3_flex_fl2va_v1_bf16.safetensors in models/loras/, the Qwen text encoder and both H3 decoders in their folders. Load your opening still and your closing still. Write the prompt in Pika's three fields: the shot itself, overall_soundscape (one to four sentences of ambience and physical sounds) and non_diegetic_music (or N/A). Keep six steps, the lcm sampler, CFG 1.0, add-on strength 1.0, and do not stack a turbo add-on on top.

How it works. The two stills pin where the shot starts and ends; the model invents the motion and the sound in between, in one continuous take unless you ask for cuts.

Why it is good. You storyboard in stills you control, then spend generation only on the in-between.

Where it breaks. Comfy's own notes: camera moves can land 0.2 to 0.3 seconds off or in the wrong direction, "wide shot" often renders as medium, and near-touches become contact. Memory needs are data-centre class, and the licence rules out local use in the US, EU, UK and South Korea.

2. Build a sound scene in layers with SFX v3. Source: ElevenLabs SFX v3 post and sound effects docs.

The steps. Start with the main action ("a heavy wooden door creaks open"). Generate, listen. Add the surroundings in the same prompt ("in steady rain, a dog barks twice in the distance"), as ElevenLabs suggests: "Start with the main action, add the surrounding sounds, and listen to how they work together." Set prompt influence high for literal results, low for variations. For a bed under a scene, turn on looping (the docs' example is soft rain). Export 48 kHz WAV for anything non-looping that goes into an edit.

How it works. One generation renders every layer together, so the rain sits behind the door instead of on top of it.

Why it is good. A podcast editor or indie filmmaker gets a placeable ambience in one pass instead of stacking four library files.

Where it breaks. The docs warn exact counts are not always followed (that "twice" may not land), 30 seconds is the per-generation ceiling, the API model is limited access by region, and a fixed duration costs 40 credits per second.

3. Audit the brand kit before you generate ads. Source: Krea Ad Studio post.

The steps. Paste the shop URL. Wait for the kit. Before generating, rewrite the selling points, fix the tone line and remove products you do not want shown. Pick a direction (product spotlight, real-world setting, editorial, or why-it-works). Generate a batch, ask Krea Agent for new headlines or sizes, preview in the Instagram and Reels mockups, download the ZIP.

How it works. Every ad pulls copy and colour from the kit, so the kit is the brief, and a small shop gets every placement size in minutes.

Where it breaks. Static ads, compute units per ad, and Krea publishes few limits beyond 12 ads per batch and 20 products per kit.

Worth testing

Pika H3 Flex Space: six-step video with sound, free. Tradeoff: shared ZeroGPU queue, and the H3 licence still governs outputs from the open weights.
ElevenLabs Sound Effects: layered scenes from one prompt. Tradeoff: credits per second when you fix duration, and counts drift.
Krea Ad Studio: a campaign's worth of sizes from a URL. Tradeoff: paid plans only, and the copy is only as good as the kit you edit.
SnapRef v1b: one character image, new poses, local. Tradeoff: non-commercial, and skin tone can drift lighter.

What actually matters from today's signal

Prompting is turning into producing. Pika's prompt guide reads like a shot list: numbered shots with cut times, speaker IDs, a field for score, a label and a role for every reference. Krea starts with research, not drawing. ElevenLabs tells you to build sound from the main action outward. These tools reward what an assistant director or sound editor already does: gather the right material and say exactly what each piece is for.

Generation gets cheaper and faster (six passes, one cent a second, a kit in under a minute) while preparation gets heavier. Budget the hour for references and the kit, not for rerolls.

The counter-signal is access. The most capable open release today needs a data-centre card and is licensed so that creators in its maker's own country cannot run it locally. Open weights you cannot legally use at home are a press release for the paid API. Read the licence before you download 211 gigabytes.

Source access notes: Quiet lab window for creators: nothing new October 28 to 30 from Adobe (latest Releasebot entry October 16), BFL, Stability, Suno, fal, Replicate, Recraft, HeyGen or Pika's blog index. Gemini 4 Carbon (October 28) has no creative features; Gemini's Connected Apps post names Adobe without saying what it does, so it is omitted. Failed or empty: OpenAI news, Hugging Face Papers, Ideogram changelog, Runway news, ByteDance Seed blog. Midjourney via WebSearch only, nothing current. HN Algolia stale; HN front page had no creative stories. Trendshift shows totals, not deltas; debpalash/VoiceStudio left out (created April, sources disagreed on stars). Civitai and blog.comfy.org not attempted. All Hugging Face dates are API createdAt values. Adversarial fact-check ran (Sonnet subagent): it confirmed the Pika dates, hardware figures, prices, licence text, Comfy settings and limits, ElevenLabs limits and Krea figures, and caught these, all fixed: an unsupported "first downloadable model" claim, finishing and stems implied inside the $0.03 price, missing combined-duration reference cap and $20M licence threshold, the 94 GiB host figure lacking Pika's page-cache caveat, a dropped word in the Comfy quote, the WAS "popping" remark being relayed by a tester, "teaser" wording on the Krea 2 add-on, an overstated "static only" for Krea, and ElevenLabs claims needing attribution to the docs rather than the blog (including the blog's "available now" against the docs' limited access). It could not confirm the Space is free to use or Trendshift placement.