Instagram story viewer> @koshimazaki> Posts

Koshi Mazaki

@koshimazaki

2.1K

followers

1286

following

Agentic AI · ComfyUI · Modular Synthesis · Building SIDKIT

POSTS STORIES REELS TAGGED

Download All

It was such a pleasure and awesome opportunity to play @cultofcryptoart opening party at @nonfungibleconference with so awesome group of artists really stoked that have joined the cult 🔺❤️🔥 good vibes killer sets from everyone 🚀

Pavilhão Carlos Lopes where the conference is taking place has awesome and really unique setup. 4x projection screens of different and custom screen sizes with 3x2m led box in the middle of the venue. It’s massive creative exploration playground! I generated visual content for each of those specifically using new AI workflows. 15min of material was created to spec especially for the event. Since AI models are trained on different resolutions this made it look even more glitchy and unique then usual which I love. I played my tunes and processed them live with modular synth treating it as an immersive experience. Was really cool to see how each artist approached it as well, as each set was so different 🙏🙌
@arthr.eth @basseadx @plutonicmind @dundomaroye @angietaylorartist

#nft #cryptoart #nfc2024 #aiart #banodoco #animatediff #eurorack #modularsynth by @koshimazaki

68

3

2 years ago

Stoked and humbled that Glitch Goddess Sweep was selected in @deforum_art AI art competition. Piece was selected among 5 out of 121 entries and will be exhibited in LA @brightmomentsgallery where 2 winners will be selected. Selection was made by AI veterans and top contributors to the community. Amazing entries so really tough competition of the most nerdy AI artists out there 🤟✨🍻♥️

Visual: SD, A1111, Multicontrolnet, custom dreambooth and Lora models, Parseq for sequencing and audio analysis. Piece is entirely generated in Google Collab

Audio: Digitone sampled in Marphagene processed by Cwejman MBC3, MMF2 and Modcan Dual Delay kicks made with BLD2🤟

#aiart #stablediffusiongui #stablediffusion #generativeart #eurorack #modularsynth #digitone #elektron #cwejman #midjourney #midjouneyart #newmedia #digitalart by @koshimazaki

72

12

3 years ago

Exited to be presenting new work Supernatural Arrival at @brightmomentsgallery in London among selected artists using @deforum_art at DREAMFORUM EU event. Massive 4K screens will be used 🙌 flip the phone to see the details ⤵️
Headphones ON 🔈✨
Here audio controls material change
Res4: wood
Piston Honda: plastic
BLD/Morphagene: metal

#eurorack #eurorackmodular #modularsynth #stablediffusion #sdxl #deforum #parseq #aiart #ai #generativeart by @koshimazaki

40

1

3 years ago

Introducing AKUSPACE — an Audio LoRA for @ltx.io 2.5 model aka the Reverb LoRA — Sound ON! 🔊🎧

I built this LoRA around a simple question:
Can a generative model learn where a sound exists?

By training on matched dry and processed versions of the same performance, AKUSPACE learns the acoustic space around the sound—its reflections, decay, ambience, and sense of distance.

AKUSPACE lets you place a voice, beat, or instrument inside an acoustic space during generation, from a small room or empty club to a cathedral or outdoor environment.

The goal is simple: make generated videos feel acoustically aware, so the sound belongs inside the scene instead of feeling added afterwards.
I built the training dataset entirely from my own material, recorded and produced over the years: Eurorack patches, drum beats, spoken voices, electronic instruments used over the years, but here model was learning how reverb responds to those sound sources. I used Eurorack modular and Ableton Live to generate training data.

For each room or effect, I created gentle, moderate, and heavy versions—moving from subtle to heavily processed. These levels are included in the training captions, making the intensity controllable through prompting.

CFG provides a second intensity “fader”: pushing it towards 4 can make the treatment louder, more detailed, and more pronounced.

I also created custom ComfyUI nodes to make prompting easier and more playful, including an interactive 3D spatial visualiser built with Three.js and rendered directly inside the node.

Interactive demo: https://akuspace.pages.dev
ComfyUI nodes: https://github.com/koshimazaki/ComfyUI-Koshi-Nodes
Model on Hugging Face: https://huggingface.co/KoshiMazaki/akuspace-ltx25
Go experiment and let me know what you make.

#LTX25 #LoRA #AudioLoRA #AKUSPACE #ReverbLoRA by @koshimazaki

46

6

a month ago

I’ve been experimenting with the @bfl.ai API this week and added FLUX 3 Video support to an API Control Surface I built. It includes the following

- Video Scripts — build keyframe batches from your asset collections: pin/vary slots, permutations, and a live job-count + cost preview before anything paid runs
- A server-owned generation queue — pause/retry, crash recovery
- Built-in model evaluation — every run captured with settings, timings and cost, rate results, export JSONL
- Full agent surface with 30 MCP tools + a CLI, so agents can plan, queue and evaluate video batches end to end
- Drag-and-drop everything, up to 10 keyframes per video
If anyone would like to try it, you can find link in comments 🫶

Feedback is very welcome!

#FLUX3 #BlackForestLabs #CreativeCohort #CreativeTools #AITools by @koshimazaki

7

0

2 months ago

Generating action-oriented content with Black Forest Labs FLUX 3 Video. @bfl.ai

I’m impressed with the model overall. It’s still evolving in early access, but the results are already stunning.
What I enjoy most is its aesthetic and world awareness. I don’t have to explain every detail, in many cases, it interprets the context correctly and adds creative choices of its own. The results remain coherent even as the speed, physics and visual language change. It feels less dependent on shot-by-shot instructions and more capable of interpreting the intent already present in the input.

It’s genuinely inspiring to work with, and being able to push it at this stage is awesome 🙏

#FLUX3 #BlackForestLabs #FLUXCreator #EarlyAccess by @koshimazaki

8

0

2 months ago

FLUX 3 finds timelines in character sheets.

I’ve been experimenting with character sheets using FLUX 3 Video through the @bfl.ai Black Forest Labs Creator Program.

The model is still evolving in early access, but I’m already getting strong first-pass results. The three examples in the video were all initial attempts across different model versions—and became personal favourites.
The video shows three animation approaches:
Fluid transitions from separate character poses into a coherent animation
Character parts interpreted as an animated model decomposition
A multi-shot sequence where FLUX 3 turned parts of the character sheet into titles
What stands out is the model’s general visual knowledge. With relatively simple prompts, it can infer motion, treat character sheets like storyboards, and make product-style shoots much more straightforward.
It also feels promising for future LoRA workflows: the model has a broad aesthetic range and shifts fluidly between very different visual worlds.
I’m still testing, but this is already changing how I think about character sheets—not simply as reference images, but as timelines for motion and storytelling.

#FLUX3 #BlackForestLabs #FLUXCreator #EarlyAccess by @koshimazaki

8

2

2 months ago

Testing new and coming FLUX 3 model with @bfl.ai creative cohort. Love how smooth output it generates even with basic prompts. This is raw generation straight from the model no post processing involved.

Will share more outputs as I go, so much to explore here. Exciting times.

#flux3 #generative #aiart by @koshimazaki

12

4

3 months ago

R3F as audiovisual Seedance 2.0 guidance now with sequencer / timeline

#aitools #threejs #generative #reactthreefiber by @koshimazaki

8

4

3 months ago

FLUX API Control Surface @bfl.ai
With MCP prompt morphing scripts and audio scheduling for video generation and much more

#aitools #claudecode #aiart #creative #generative by @koshimazaki

6

5

3 months ago

I’ve been turning the React Three Fiber / Three.js experiment into an audio-reactive reference video generation setup.

The goal is to control multimodal generation with audio, generated motion as a control layer, and eventually model fine-tuning based on the outputs.

Instead of generating a video from a prompt and adding music later, the audio becomes the structure of the scene and drives the generation.

The setup can now:

• slice audio into up to 15-second loops
• sync/guide 3D / React Three Fiber motion to the track
• place timeline markers for beats, chapters, and visual transitions
• turn markers into prompts with shot duration information
• layer multiple camera movement tracks that morph into new animations
• capture the reactive 3D world in different aspect ratios
• export MP4s with the music embedded
• use those exports as guidance/reference videos for AI video generation

So the browser scene becomes a kind of controllable world-building layer.

Objects, camera moves, rhythm, cuts, and spatial changes can all be composed together before sending anything into multimodal generation.
The model responses are not fully consistent yet, so more experiments are needed. But as a bulk generation workflow, it already looks useful for editing, especially because it can produce unusual motion that would be difficult to achieve otherwise.

The clip below shows the scene setup, audio-reactive reference generation, and the resulting output, cut into an edit.

Thanks @magnific_ai for the compute 🙏

Images used in the process are generated with FLUX from @bfl.ai Black Forest Labs and GPT Image 2 from @openai

#AItools #FLUX #Seedance #ThreeJS #ReactThreeFiber by @koshimazaki

11

2

3 months ago

Experimenting with React Three Fiber / Three.js as a new pipeline for audio-reactive video generation.

The idea is to use as a controllable world-building layer with primitives, GLB assets, camera movement, object placement, and audio-reactive motion all living inside a scene editor in the browser.

From there, I can export audio-visual mp4 files, guidance/reference frames, and sequence them with @bfl.ai Black Forest Labs FLUX.2 images, and use them to drive Seedance 2.0 video generation. Later idea is to train or fine-tune video or multimodal models.

What’s interesting here is the control. Instead of treating audio as media added after the video is edited, audio becomes integral part of the generative process. Rhythm, motion, cuts, spatial changes, and visual actions can all be composed together.

In theory, the same setup can drive realistic video, pseudo-realistic worlds, anime-style scenes, or abstract audio-reactive visuals.

Still early days, but it feels promising. Thanks @magnific_ai for the compute 🙏

#ai #generative #pipeline #aitools #threejs by @koshimazaki

8

2

4 months ago