🎥 New Open-Source Minimax H3 Fork

FreeVideo is a local inference engine for MiniMax H3 on consumer GPUs, built on OpenVDN's 8-step VDN-H3 model with Video DeltaNet's hybrid attention.

It coordinates VRAM, system memory, and disk, adapting weight placement, compute precision, and attention kernels to the available hardware.

FreeVideo runs as a ComfyUI plugin, with a Windows launcher for setup and command-line support on Linux. Its core features include:

⚪️Hardware-adaptive execution: Chooses the FP8 compute path for each GPU architecture, either native FP8 or FP8 storage with BF16 compute, and automatically probes the available attention kernels.

⚪️Low-memory inference: Weight streaming, asynchronous prefetching, and chunked computation keep peak memory low, enabling inference with as little as 8 GB of VRAM and 16 GB of RAM.

⚪️Multimodal inputs: Text prompts, first and last frames, and image, video, and audio references.

⚪️Community LoRAs: Use MiniMax H3 LoRAs in your workflow. See examples.

⚪️ComfyUI integration: A dedicated creative workspace inside ComfyUI that supports two-pass sampling and batch generation and keeps a history of past creations. For finer control, switch to the node view to add LoRAs or customize the workflow.

⚪️One-click deployment: The Windows launcher sets up ComfyUI, the runtime environment, and the models, reuses existing models, and supports offline installation.

How to install on Windows and everything else about this model, see here https://github.com/FlashML-org/FreeVideo

🔤 AI News Roundup #2

Dive into the latest developments in the AI world. Leave an emote if you want more recaps like these.

🖱️Ideogram v4.5 - Has been released with impressive multi-turn edits that do not turn images to mush after multiple revisions. [Review]

🖱️Alibaba Mystery Model - Alibaba teased a new video model at the Apsara event. It is unclear whether this is Wan 4.0 or Happyhorse 1.2, but more information is expected in October/November. This model can generate clips of at least 30 seconds in length.

🖱️Kling 4.0 Flash Launch - Now available exclusively on Kling's own platform. A standard Kling 4.0 release is also expected later, along with an API release for Flash soon.

🖱️Gemini 4 Argon - Google announced a new model with no current release date. Benchmark tests show better results in certain categories compared to Astra and Opus 5.5.

🖱️ElevenLabs v4 Release - A brand-new, highly expressive, and powerful model that demonstrates a very wide array of tonal range.

🍣Sushi Construction (Minimax H3)

- Reference(s): No
- Styles Support: Realism

"[Shot 1] Photoreal cinematic macro miniature photography, extreme close-up.
A miniature Japanese construction crew is building an enormous piece of salmon nigiri as if it were a skyscraper. Tiny workers in orange safety uniforms and yellow hard hats climb bamboo scaffolding attached to a giant block of white sushi rice. A massive glossy slice of salmon hangs above them from a miniature crane. The scene is lit with warm restaurant lighting, with gentle steam drifting through the background and shallow depth of field creating a tilt-shift miniature effect.
The workers guide the suspended salmon into place using ropes and hand signals. Small forklifts transport giant rice grains across the wooden sushi counter. Tiny ladders, ropes, cones and miniature construction equipment are scattered around the site.
The camera begins in an extreme macro close-up on one worker standing beside a single rice grain, then slowly pulls backward with medium amplitude at slow speed while arcing slightly to the right, gradually revealing more of the giant sushi construction site.

[Shot 2] At 00:03.500, the camera cuts to a lower side angle medium close shot.
A small crane lowers the translucent salmon slice into alignment on top of the rice block while several workers on the scaffolding steady it with guide ropes. One worker waves signaling flags while another directs a forklift carrying wasabi like a construction material. Steam curls upward behind the structure.
The camera trucks left with medium amplitude at moderate speed, keeping the descending salmon and workers centered as the scale difference becomes more obvious.

[Shot 3] At 00:07.000, the camera cuts to a wider reveal shot.
The full sushi construction site is now visible on the wooden counter: scaffolding around the nigiri, multiple tiny workers, cranes, forklifts, and a miniature work zone surrounding the completed salmon topping. The crew continues moving with purposeful activity while warm restaurant bokeh glows in the background.
The camera slowly pulls farther back with small amplitude at slow speed and settles into a final heroic wide shot of the completed sushi skyscraper and the bustling miniature crew.
overall_soundscape:
Soft restaurant ambience, distant dish sounds, miniature forklift motors, crane winch sounds, footsteps on wood, light tool clinks, rope tension sounds, subtle steam hiss, workers giving short shouted instructions.
non_diegetic_music:
Playful whimsical orchestral music with pizzicato strings, light percussion and subtle woodwinds. The music gradually builds as the construction progresses and ends with a soft satisfying fade out."

✨ Get access to Minimax H3 at OpenArt
(20% discount /w code "L1TMGOQ2" from all plans)

🌐 Google has released Nano Banana 2.1 on Flow; I'm not sure if it is available anywhere else right now. I will keep you posted, as it definitely seems to be an improvement over the older model.

But it is not going to replace Ideogram or GPT Sunburst/Flare for sure.

Note: There might be region restrictions on the model atm or its been pulled out?

🚌 Origami Bus (Wan 3.0)

- Reference(s): No
- Styles Support: Realism

"15 seconds, 9:16, 25 FPS. Three connected cinematic shots. A mustard-yellow city bus is stuck in a daytime traffic jam. Keep its color, windows, passengers and driver consistent throughout. Photoreal city with one tactile, impossible transformation: the bus folds like a single sheet of thick paper while everyone remains safely inside.

00–03s: Street-level three-quarter view. Cars are motionless. The bus driver taps the steering wheel; a passenger glances at the long line ahead. A crisp fold suddenly forms across the bus roof.

03–07s: Stay with the bus as its body folds in a precise origami sequence. The wheels tuck inward, the sides crease and the entire bus becomes one mustard-yellow paper glider. The passengers remain visible through its folded window shapes. Nothing breaks off.

07–11s: Cut to a smooth tracking shot as the glider sails just above the traffic. Its moving shadow passes over the stationary cars. Follow its graceful descent toward an empty bus stop beyond the jam.

11–15s: The glider touches down and unfolds into the same full-size bus. Its doors hiss open. One passenger steps out, notices a sharp paper crease in their coat, pats it flat and walks away as though this were routine. Hold briefly on the bus.

Audio: idling traffic, rhythmic paper-folding sounds, a short rush of wind, soft landing, bus-door hiss. No dialogue or music. Clear spatial continuity; no duplicate bus, disappearing passengers, logos, captions or text."

✨ Get access to Wan 3.0 at OpenArt
(20% discount /w code "L1TMGOQ2" from all plans)

✨Vidu Q4 Model Release

Vidu has released a new flagship AI model that includes more nuanced facial expressions, smoother cuts, dynamic camera movement, and action continuity. Among its features:

- Supports up to 15 image references
- Supports up to 3 audio references
- 2K/4K output with 10-bit color depth
- Pricing at $0.014 per second (Cheap)
- Up to 16 seconds of duration

Right now, it is available on Vidu's own platform. I also saw WaveSpeed post that it is available there as well.

🔤 AI News Roundup #3

Dive into the latest developments in the AI world. Leave an emote if you want more recaps like these.

🖱️Grok Imagine Video v1.5 Lite - xAI released this new model at a very affordable cost, starting at $0.02 per second for 480p. [Review]

🖱️Nano Banana 2.1 Out - Google has updated their image model with improved text and accuracy. [Review]

🖱️Mistral Large 4 Preview - The model dubbed "le chonk" has been showcased, said to be the best open-source LLM, beating its Chinese counterparts.

🖱️OpenAI Retiring Custom GPTs - OpenAI has announced it will retire Custom GPTs and move the specialized workflows built around them to a new plugin architecture spanning ChatGPT and its Codex coding agent.

😀 African Matrix (Wan 3.0)

- Reference(s): No
- Styles Support: Realism

"Photorealistic Zimbabwean underground parking garage at night. Real actors, real environment, practical-looking impacts, premium cinematic VFX. Takunda is always the same person: 160cm, small build, African tank top. Natural dark skin, expressions and realistic movement. No CGI human or digital double.
The Agent is a calm Somali man in a perfect red suit, white shirt, black tie and white sunglasses. Short slicked hair, emotionless, precise and unarmed.
Style: classic bullet-time / wire-fu action. Real Wing Chun, Taekwondo and Jeet Kune Do. Fast combat with brief extreme slow-motion moments and smooth 180°/360° camera orbits. No visible wires, floating, glowing energy, auras or colored effects. All impacts obey real physics.
Sound: No music or subtitles. African dialogue only. Echoing footsteps, breathing, fabric movement, punches, impacts, fluorescent buzz and garage ambience.
TIMELINE
00:00–00:03

Takunda stands still in the empty garage. The Agent watches from 20 feet away.

Agent: “Waan ku daawanayay.”

(“I've been watching you.”)
00:03–00:06

The Agent suddenly rushes forward and punches at Takunda’s face. Takunda instinctively tilts his head; the fist barely misses.
00:06–00:10 — BULLET TIME #1

Extreme slow motion. Camera orbits 180° around the missed punch. Takunda’s tank top and short hair react to the passing fist; dust hangs in the air.

TIME SNAP: Takunda counters with an elbow. The Agent blocks effortlessly.
00:10–00:14

Rapid Wing Chun exchange: traps, parries and redirects. Neither lands a clean hit.

Agent: “Aad bay u daahday.”

(“You woke up too late.”)
00:14–00:18

The Agent strikes Takunda’s shoulder, sending him into a concrete pillar. Dust bursts from the impact. Takunda rebounds into a spinning back kick.
00:18–00:22 — BULLET TIME #2

360° orbit around the connecting kick. The Agent’s suit and sunglasses move in slow motion; broken fluorescent glass hangs in the air.

TIME SNAP: The Agent crashes onto a parked car, visibly denting the hood.
00:22–00:26

The Agent immediately gets up, fixes his tie and attacks again. Fast parries, redirects and a low sweep. Both move with extreme precision.
00:26–00:29

Brief micro slow-motion. Takunda narrowly dodges, redirects the Agent’s arm and lands one powerful palm strike to his chest. No energy effect—only physical impact.
00:29–00:30

TIME SNAP. The Agent skids across the floor. His sunglasses fly off. Takunda stares at his own hand, shocked and breathing hard. The Agent lies still. Takunda walks away into the flickering light without looking back.
Camera: Kinetic low-angle tracking during normal speed; smooth orbital bullet-time during key moments; whip-pans and impact shake only on contact.
Core rule: The spectacle comes entirely from realistic martial arts, extreme reflexes, physical impacts and bullet-time cinematography—not supernatural energy."

✨ Get access to Wan 3.0 at OpenArt
(20% discount /w code "L1TMGOQ2" from all plans)

🥳 Apparently, Bytedance is already working on a new variation of a model. The question is: Does FLASH mean FAST?

If so, that would mean you can now create a lot more affordably good dialogue scenes. It matches the release schedule, so I guess this would drop this month or in November.

😭 Messy Hair (Seedance 2.0)

- Reference(s): No
- Styles Support: 3D

"Create a cute, polished cinematic 3D animated video, exactly 15 seconds long, using the attached character sheet as the strict visual reference.
CHARACTER
A cute young Black girl with warm medium-brown skin, large expressive brown eyes, soft rounded facial features, and very thick, dense, tightly defined red curls with lots of volume and individual coils. She wears the same white long-sleeve top and light lounge pants from the reference.
Keep her face, skin tone, body proportions, outfit, red hair color, and character design identical throughout the entire video.
SCENE
She sits directly in front of her bedroom vanity mirror in a cozy pastel bedroom. Warm soft lighting, cute feminine decor, polished stylized 3D animation, expressive facial acting, cinematic depth of field.
15-SECOND ACTION
0–3 sec :
She looks into the mirror and quickly but carefully brushes her huge curly hair section by section. She gathers the brushed sections toward the back of her head.
3–7 sec :
She gathers all her hair upward and twists it into a high bun. She uses the brush to smooth the front hairline and side sections backward, making those areas sleek and neat while the gathered hair remains visibly curly and voluminous.
7–9 sec :
Just as the bun looks perfect, one stubborn curly strand springs out from the side.
She notices it in the mirror and gives an annoyed little look.
She quickly brushes the strand backward, tucks it into the bun, and secures the bun with several small hairpins.
9–11 sec :
She checks herself in the mirror and smiles proudly, satisfied that the bun is finally secure.
11–15 sec :
POP! POP! POP!
The hairpins fly loose and the bun completely unravels.
Her entire mass of thick, dense red curls instantly falls back down around her shoulders, returning to the huge curly hairstyle from the beginning.
She freezes, stares at herself in the mirror, and gives an exaggerated annoyed, defeated expression.
End on her frustrated face for a comedic beat.
VISUAL & MOTION
Smooth high-quality 3D animation, cute exaggerated expressions, natural hand movements, realistic curly-hair physics, bouncy individual curls, subtle camera movement, soft cinematic lighting, and polished character animation.
Use quick but readable cuts between the vanity medium shot, brushing close-up, bun close-up, stray curl, hairpins, and final reaction.
IMPORTANT: The loose hair must be extremely curly and dense, not wavy or straight. The front and side hair becomes sleek only while she intentionally brushes it backward for the bun. When the bun breaks, all the hair returns to its original large, tightly curled, voluminous state.
No text, subtitles, logos, watermark, extra characters, or changes to her appearance."

✨ Get access to Seedance 2.0 at OpenArt
(20% discount /w code "L1TMGOQ2" from all plans)

🛠 (REA) = Reverse Engineer Anything

For those who want to jump into reverse engineering, this is the tool for it: (https://github.com/morluto/rea)

REA connects your agent to tools for inspecting native binaries, JavaScript and Electron apps, .NET assemblies, and websites. You can also use the same tools from your terminal. Analysis runs locally, and results include the evidence and limitations behind each conclusion.

Setup registers REA with your agent and installs matching workflow instructions. Native analysis can use an existing Hopper, Ghidra, or IDA installation; setup can optionally install Hopper with approval. Static JavaScript analysis needs no native analysis engine.

This is available for Linux/Windows. Supports Claude Code, Codex, Cursor, Gemini CLI, Grok build and countless other agents.

🎆 1000 Subscribers Reached!

Before anyone trolls and unsubscribes, I just want to say a big thank you to all the followers here. I will be posting more prompts and news in the coming days.

🎁 Plus, if you stay subscribed, you will be entered into a raffle for a FREE Telegram Premium subscription, which we will give away soon to celebrate this monetization milestone.

➡️I also want to feature a AI contest tool that member of our community created in our discord server - https://entr.win/

👨‍💼 Swapping Outfits (Seedance 2.5)

- Reference(s): No
- Styles Support: Realism

"Each frame can change his suit independently. Do not maintain outfit continuity during the walk. Ultra-realistic cinematic video, 15s, 24fps. The video starts mid-motion: a strikingly handsome man (chiseled jawline, strong brows, sharp cheekbones, styled dark hair, tall athletic broad-shouldered build, confident alpha "Chad" energy) is already striding forward through a long, elegant corridor. No intro, no standing start. His walk is smooth, natural & continuous throughout.

SETTING: Luxurious hotel-style corridor, polished marble floor, warm wall sconces, dark wood paneling, soft depth haze. A tall full-length mirror with an ornate gold frame hangs on one wall near the end of the corridor.

CAMERA (0–12s): Front-facing tracking shot, camera moving backward ahead of him, slight handheld energy, fast push-ins & subtle orbits matching his walking rhythm. Full body visible, centered in frame.

OUTFIT TRANSFORMATION (0–12s, VERY HIGH SPEED): His suit changes extremely fast, near every-frame flicker, synced to his footsteps. Every frame = new suit, no repetition, no blending:

- classic navy two-piece
- charcoal pinstripe three-piece
- black tuxedo with bow tie
- white dinner jacket
- burgundy velvet blazer
- grey double-breasted chalk-stripe
- light beige linen suit
- emerald green tailored suit
- tweed check suit with waistcoat
- all-black slim suit with black shirt
- royal blue suit with open collar
- brown corduroy suit
- cream suit with turtleneck
- metallic silver evening suit
- Glen plaid suit

Shirts, ties, pocket squares, shoes & accessories (watches, tie clips, sunglasses) also change with every flicker.

FINAL SHOT (12–15s): Hard cut to a side angle, framing him in profile as he reaches the mirror, with both him & his reflection visible. The outfit flicker stops & settles on his final look: an impeccably tailored midnight-blue three-piece suit, crisp white shirt, silk tie, polished black oxford shoes. He slows slightly, turns his head toward the mirror, meets his reflection & gives a confident, satisfied smile, as if he has finally chosen this suit. Then he continues walking & exits the frame, leaving the mirror reflecting the empty corridor.

CRITICAL: His face, hairstyle, body proportions & identity stay 100% consistent across every outfit change. Only the clothing changes, NOT the person. His mirror reflection must match him exactly.

SYNC: Outfit changes lock tightly to his footsteps & camera motion.

VISUAL FEEL: Hyper-fast flicker, energetic but clean & readable, luxury fashion-commercial aesthetic. No distortion or warping of body or face.

LIGHTING: Warm cinematic corridor lighting, soft highlights on fabric textures, subtle reflections on the marble floor.

CONSTRAINTS:

- walking motion must remain perfectly continuous through all outfit changes

- no pose resets, no stopping until the mirror moment
- no face distortion or identity change
- flicker ends exactly at the cut to the side angle
- no text, no logos"

✨ Get access to Seedance 2.5 at OpenArt
(20% discount /w code "L1TMGOQ2" from all plans)

🟢For Those Who Create. KLING 4.0 Is Now Fully Available.

What’s New, Starting Today! Access Kling 4.0 through a refreshed creation interface on kling.ai Access Kling 4.0 through MCP and CLI for greater flexibility in your video creation workflow https://kling.ai/h5-app/invitation?code=7BY9XRUN6H9T

Coming Soon!
➡️10-bit HDR video output at 4K and 1080p resolutions API access for enterprise customers and developers.
➡️On October 21th the API release happens, thats when you are going to see 3rd party sites have it, for now you have to subscribe to kling to use it.

You call the shots. Let the creating begin

🥊 Alien Boxing Match (Kling 4.0)

- Reference(s): Yes (First/last frame + multikey frames)
- Styles Support: Realism (tested)

"19-second photorealistic cinematic science-fiction fight sequence. Use the seven supplied images as key visual anchors, preserving both fighters’ identities, proportions, clothing, blunt armored gloves, and the underground cage arena. Begin with the wide face-off and end with the human on one knee, looking defiantly up at the alien.The human fights with speed, tight boxing technique and evasive footwork; the larger alien uses heavy, forceful punches and relentless forward pressure. They circle briefly before exchanging fast, committed strikes. Cut to an intense close-up of the human behind his raised gloves, sweat running down his face. The alien lands a powerful punch: enter brief extreme slow motion as the glove compresses the human’s cheek, his head turns, and small blood droplets and sweat scatter through the light. Return sharply to real-time action.An overhead shot captures the human ducking beneath a wide hook and stepping inside the alien’s reach. Cut low as he drives a compact counterpunch into the alien’s ribs, making its torso recoil. The alien recovers and forces him against the cage. In a tight, slightly canted shot, the human blocks a heavy punch with both gloves; the steel mesh shudders behind him. A final forceful exchange sends him stumbling away and dropping to one knee. End at floor level: one glove planted in the dust, the other protecting his face, eyes fixed on the exhausted alien still standing across from him.Use deliberate cinematic cuts between the reference angles. Connect each moment with convincing fight choreography, clear punch contact, defensive reactions, realistic balance and weight transfer. Keep the action moving between the keyframes rather than pausing on them. Reserve slow motion for the main facial impact; all other combat is fast and grounded.Soft industrial overhead lighting, muted amber accents, restrained contrast, realistic skin, shallow depth of field in close-ups, clean image without grain. Heavy glove impacts, strained breathing, boots scraping grit, cage rattles and muffled crowd reactions. No dialogue, no music, no sharp weapons, no glowing powers."

✨ Get access to Kling 4.0 access now!