Skip to content

Is Veo 3.1 worth the attention? Honest review and results

2025-10-15 · livestream / screening · Ruslan · recording, 1 h 46 min · watch on YouTube

Topic: Video generation & shot direction (also: Prompting & AI assistants, Assets, references & consistency, Failure modes & fixes, Higgsfield platform & ecosystem, Audio, voice & music, Cost, credits & plans) · mode: Review & comparison · tags: time-bound offer, giveaway, prompts shared, one creator's experience

In a livestream on the day Higgsfield released Veo 3.1, Ruslan tests the model live while it is unlimited for a week on the Creator and Ultimate plans. He shares working habits: timeline prompts by seconds, music set to false in JSON prompts, no capitals in dialogue, and ChatGPT to rewrite refused prompts; clips run only 4, 6 or 8 seconds. Scene cuts work on slow scenes, so he sees product ads, UGC or cinematic, as its good use, while photos of real people such as Elon Musk were blocked. His verdict is a slight evolution over Veo 3.0, with no better likeness and Sora 2 still a little ahead.

What is said

The parts of the video and what is said in each (★ — how much it matters to the video).

00:00:00 — Ruslan opens the second stream: Veo 3.1 released on Higgsfield, an Unlimited week, promo codes and the test plan

  • ★ Ruslan introduces himself as head of the prompt engineering team at Higgsfield, hosting its second stream. — 00:00:24 · Higgsfield AI
  • ★★★ Ruslan announces that Higgsfield has just released Veo 3.1 on its website and platform. — 00:00:24 · Veo 3.1, Higgsfield AI
  • ★★★ Ruslan: Higgsfield makes Veo 3.1 unlimited on the Creator and Ultimate plans for one week, free to use without spending credits. — 00:01:39 · Veo 3.1, Higgsfield Unlimited
  • ★★ Ruslan: From the release notes, Ruslan says Veo 3.1 takes up to three image references (a product, yourself, a scene or a style) as well as start and end frames. — 00:04:08 · Veo 3.1, keyframe
  • ★★ Ruslan says Veo 3 and Veo 3.1 have their own reasoning models behind the video model, so simple prompts work. — 00:04:08 · Veo 3, Veo 3.1

00:05:01 — First tests: a selfie speaking Japanese, a singer, an Ingredients scene; Veo 3.1 Fast versus Veo 3.1, black borders, sound

  • ★★ Ruslan: Veo 3.1 on Higgsfield offers a 9:16 aspect ratio, an 8-second duration and 720p quality among its settings. — 00:05:58 · Veo 3.1
  • ★★ Ruslan: Veo 3.1 Fast generates faster at somewhat lower quality; the standard Veo 3.1 gives higher-quality, crisper video. — 00:06:27 · Veo 3.1
  • ★★ Ruslan says the major problem of Veo 3 was its sound, heavily compressed and full of artifacts. — 00:07:28 · Veo 3
  • ★★ Ruslan warns that a start image that is not 16:9 or 9:16 will most probably give Veo 3.1 clips black borders. (hedged) — 00:09:16 · Veo 3.1, keyframe
  • ★★ Ruslan: In his singer test, Ruslan hears that Veo 3.1's sound quality has improved. — 00:10:02 · Veo 3.1

00:10:40 — A JSON prompt schema built in ChatGPT; a watch close-up keyframe in Soul and Seedream; the Sora 2 prompter GPTs

  • ★★ Ruslan's JSON prompt schema, written by a ChatGPT prompt enhancer, lists frame rate, film grain, subject, wardrobe, scene, location, time of day, action, props and physics. — 00:13:03 · ChatGPT
  • ★★★ Ruslan: In his JSON video prompts Ruslan sets music to false unless he asks for music, so that no music drops in unwanted. — 00:13:31 · native audio
  • ★★ Ruslan's JSON video prompts always exclude subtitles. — 00:14:08
  • ★ Ruslan says he copied the JSON prompt schema from a prompt engineer who posts JSON prompting and fine-tuned it to work better. — 00:14:08
  • ★★ Ruslan: In Ruslan's first Veo 3.1 Ingredients test, he came out as a different-looking Asian man, while the hot-sauce jar and the woman stayed close to the input images. — 00:15:56 · Veo 3.1
  • ★★ Ruslan: With Ingredients (reference images) selected, Veo 3.1 on Higgsfield generates only 16:9, not vertical video. — 00:16:24 · Veo 3.1
  • ★★★ Ruslan has learned that Veo 3 and Veo 3.1 work very well with structured prompts carrying a timeline: what happens in seconds one to two, two to four, and so on. — 00:22:00 · Veo 3, Veo 3.1

00:22:26 — Timeline prompts, clip lengths, Veo 3 Fast compared, Sora 2 prompts that work in Veo 3.1

  • ★★★ Ruslan says Veo 3.1 gives clips of only 4, 6 or 8 seconds; rumours that it handles 30-second single videos are not true. — 00:24:15 · Veo 3.1
  • ★★ Ruslan: After a watch close-up test, Ruslan says prompts from his Sora 2 prompt GPTs look very good in Veo 3.1, a new prompting strategy. — 00:26:49 · Veo 3.1, Sora 2

00:27:35 — Ingredients with a person, a room and a Nano Banana hoodie; likeness; dialogue without capitals; stream outages

  • ★ Ruslan: Asked about a music video for Kion (as heard), Ruslan says Kion 1 was a workflow of several video models and did not include Veo 3. — 00:32:09 · Veo 3
  • ★★ Ruslan says a couple of iterations is sometimes all a Veo 3.1 clip needs, especially with Unlimited. — 00:37:59 · Higgsfield Unlimited, Veo 3.1
  • ★★★ Ruslan advises never writing dialogue lines in capitals: Veo 3 reads a capitalised word as an abbreviation and spells it out letter by letter. — 00:40:22 · Veo 3

00:41:51 — Scene cuts in a hand-cream commercial; fast rap speech; guideline refusals

  • ★★ Ruslan: To make a character speak fast in Veo 3.1, Ruslan advises leaving out all punctuation marks, not even full stops. — 00:49:29 · Veo 3.1
  • ★★ Ruslan: Veo 3.1 refused Ruslan's rap prompt as against its guidelines; he thinks it recognised the lyrics of Eminem's "Rap God". (hedged) — 00:51:29 · Veo 3.1
  • ★★ Ruslan: When Veo 3.1 refused a hand-cream commercial, Ruslan suspects the model thought the woman's face looked like a celebrity, and swaps the woman. (hedged) — 00:55:12 · Veo 3.1

00:55:42 — Car-chase and anime scene cuts; rewriting prompts against NSFW filters; scene cuts work on slow scenes and product ads

  • ★★★ Ruslan: From a two-image hand-cream commercial with four cuts and a voice-over, Ruslan confirms Veo 3.1 makes scene cuts with Ingredients and handles the text well even at 720p. — 01:01:14 · Veo 3.1
  • ★ Ruslan finds Veo 3.1's content guidelines very annoying for now, after an anime fight prompt is refused. — 01:04:08 · Veo 3.1
  • ★★★ Ruslan: When Veo 3.1 refuses a prompt, Ruslan opens a new ChatGPT window and asks it to adjust the prompt so it does not trigger NSFW filters even slightly. — 01:05:21 · ChatGPT, Veo 3.1
  • ★★ Ruslan finds that four scene cuts work well in Veo 3.1, but he does not like its sound. — 01:09:06 · Veo 3.1
  • ★★★ Ruslan's core learning: Veo 3.1 scene cuts work well, but on slow scenes that are not action. — 01:09:34 · Veo 3.1
  • ★★★ Ruslan says Veo 3.1's good use case is product ads, either UGC style or cinematic. — 01:10:03 · Veo 3.1

01:10:34 — Celebrity likeness blocked; UGC protein-snack ads

  • ★★★ Ruslan: Veo 3.1 Ingredients refused photos of Elon Musk and another real man even with a prompt that did not name them; Ruslan says the images triggered the block. — 01:12:17 · Veo 3.1
  • ★★ Ruslan: Asked for the JSON prompt schema, Ruslan says it will be put into the description of the stream recording. — 01:13:18

01:20:19 — Q&A: Veo 3.1 against Veo 3.0 and Sora 2, the Sora 2 Enhancer, credits for refused generations, faces and brands

  • ★★★ Ruslan says Veo 3.1 improves on Veo 3.0 in scene cuts (still short of Sora), sound quality and Ingredients, which take several images without start and end frames. — 01:20:49 · Veo 3.1, Veo 3, Sora
  • ★★★ Ruslan says Sora 2 does a little better than Veo 3.1 for now, especially in scene cuts, pacing and continuity; its one drawback is flickering. — 01:21:36 · Sora 2, Veo 3.1
  • ★★ Ruslan: Higgsfield's Sora 2 Enhancer (as heard) takes a Sora 2 video with boiling and flickering and deflickers it. — 01:22:21 · Sora 2
  • ★★★ Ruslan: Asked whether Higgsfield takes credits for generations cancelled by the guidelines, Ruslan says no. — 01:23:15 · credit
  • ★ Ruslan: Asked for trained characters selectable in every video model and an ElevenLabs integration for trained voices, Ruslan says Higgsfield will definitely look at that. — 01:23:48 · ElevenLabs

01:24:13 — A Red Bull and handbag ad in Paris, the rap test; takeaways on likeness and sound; rap music videos

  • ★★ Ruslan says a face or a brand can be used in photo references, as long as the face does not look like a prominent person or celebrity. — 01:24:13 · Veo 3.1
  • ★★★ Ruslan's takeaway from a Paris handbag-and-Red-Bull ad: Veo 3.1 did not improve in character likeness, its sounds are cheap, and Sora 2 does better. — 01:30:35 · Veo 3.1, Sora 2
  • ★★ Ruslan says Veo 3.1 upgraded the speech sound quality since Veo 3, but he thinks the overall sounds got worse. — 01:31:07 · Veo 3, Veo 3.1
  • ★★ Ruslan says a rap music video can be made with Veo 3.1, but you would add your own music, a very complicated workflow: stitching the clips to the song is hard, though something short could work. — 01:31:59 · Veo 3.1

01:32:51 — Q&A: prompt enhancers, hiring, the model race, using your own voice, the GPTs and JSON prompting compared

  • ★★ Ruslan says Higgsfield will make prompt enhancers for Veo 3.1, so users can turn the enhancer on and write very short prompts. — 01:33:47 · Veo 3.1
  • ★★ Ruslan says Higgsfield is hiring many people, prompt engineers among them. — 01:34:19 · Higgsfield AI
  • ★★ Ruslan: Higgsfield also needs back-end and front-end developers, ML researchers and engineers, and marketers. — 01:34:51 · Higgsfield AI
  • ★★ Ruslan: Calling the video-model race very fast, Ruslan expects Google to come up with other models in maybe one or two months. (hedged) — 01:35:47 · Google
  • ★★★ Ruslan calls Veo 3.1 just a slight evolution over Veo 3.0. — 01:36:28 · Veo 3.1, Veo 3
  • ★★★ Ruslan: To use your own voice in Sora 2 or Veo 3.1 clips, Ruslan generates the clip, extracts its audio, separates the voice, and trains his own voice in ElevenLabs. — 01:37:51 · Sora 2, Veo 3.1, ElevenLabs, voice cloning
  • ★★ Ruslan: The HF Shot List Prompter, made for Sora 2, writes only shot lists and handles scene cuts; whatever the prompt leaves out is up to the model. — 01:40:46 · Sora 2
  • ★★ Ruslan suggests rewriting a Detailed Prompt GPT prompt to integrate a shot list, mixing the two prompt styles. — 01:41:12
  • ★★ Ruslan says JSON prompting gives two or three words per category, so the model understands more of the sequence than from plain text. — 01:42:36

01:43:06 — Promo codes, Discord support, sign-off

Threads in this video

Where the stories that moved here went next

Mentioned

Central: Ruslan · Sora 2 · Veo 3.1

Also: People: Elon Musk · Eminem

Works: Rap God (song)

Organizations: Discord · ElevenLabs · Google · Prada · Red Bull

Concepts: keyframe

Products: ChatGPT · Creator plan · Detailed Prompt GPT · Enhancer · HF Shot List Prompter · Higgsfield Unlimited · Nano Banana · Seedream · Sora 2 Enhancer · Soul · Ultimate plan · Veo 3

Also centred on Ruslan: Master Sora 2 for AI Commercials And Grow Audience

← Previous: Higgsfield Upscale is Closing the Gap Between Phones & Pro Cameras · Next → How We Created Higgsfield Popcorn? An AI Tool That Allows You to Replace Faces in a Movie Scene