Elliot Blanchard

🎉 Jellyfish AI Impact now supports Windsurf, a leader in AI-powered coding solutions, offering tools like the Windsurf Editor that blend collaborative, continuous AI assistance with developers' workflows. Together, Windsurf and Jellyfish are changing how AI and humans work together.  Through our robust integration with Windsurf, Jellyfish AI Impact allows joint customers to:  🤖 Measure Windsurf’s impact on overall developer productivity: Analyze how Windsurf solutions like Windsurf Editor impact different engineering workflows and use cases – whether it's building new features, reducing technical debt, or reviewing for code quality and security improvements. 📈 Scale Windsurf adoption across the organization: Identify teams with high engagement to further accelerate development, automate testing, and streamline deployment, then replicate these best practices and implementation strategies company-wide. Learn more about what this integration means for your engineering organization here: https://lnkd.in/ghNguE47

54

Introducing Atmospheric Search. Mood boards aren’t archives. They’re active tools. Atmospheric Search helps designers navigate visual history using the language of art direction, tone, treatment, process, and cultural context, turning stored references into usable direction.

1

2 Comments

With the launch of Luma, Multidimensional Creative Specialist Tony Hou's 3DxAI workflow lets him quickly create & control advanced effects without complex volume or fluid sims.

59

3 Comments

Atmospheric Search is here! Our team sat down with designers from some of the top studios in the world. One consistent issue we heard was that during creative concepting, reference images will get used once and then disappear in to the file abyss. Atmospheric Search allows you to search through your reference image library by categories like aesthetic signal, systemic implication, emotional register, and cultural reference so that previous reference content can be easily located and included in new creative concepts.

8

Beyond the Slop #40 — Crimson Desert AI Backlash Crimson Desert devs had to apologize and remove generative AI art from the game. Here’s the thing. It was just a piece of art on a wall. -It didn’t affect gameplay -It didn’t change systems -It didn’t replace core creative work And yet it turned into a controversy. This feels like the wrong battle. Not every use of AI is a threat to creativity. Sometimes it’s just… a background detail. If it doesn’t impact the game, removing it doesn’t improve the game either. https://lnkd.in/eUEqKBa4

2

Runway just solved one of the biggest headaches in AI video: character consistency. If you’ve ever used other GenAI tools, you know how frustrating it is when your character’s face changes with every render. The new “References” feature in Runway finally lets you lock in your character’s identity—even across multiple shots. But there’s a catch: it’s not one-click magic. The feature only works through a two-step process: 1. Generate a consistent still using Text-to-Image + References 2. Animate that still using Image-to-Video The key? Aligning your reference images with prompts that speak the same cinematic language. Think: lighting, emotion, wardrobe, lens choice. Not just keywords. In my latest breakdown, I walk through the full workflow, including: • How to prep your reference images for best results • Prompting tips that actually match character and tone • Why Runway’s “floaty” animation style can be a creative advantage • How to plan multi-shot sequences without breaking continuity Watch the full tutorial here: https://linktw.in/yPRklL This is a must-know if you’re building branded content, narrative shorts, or stylized visual experiments using GenAI. #RunwayML #AIVideo #GenerativeAI #AIContentCreation #PostProduction #ContentStrategy #CreativeTech #VideoEditingTools

2

Animating non-human characters with AI still presents real tradeoffs. I’ve been testing a hybrid workflow combining Runway (Act-One), Luma Labs, Runway Retexture, and image-to-video tools. It’s not perfect—but it’s the most promising pipeline I’ve found so far. The problem: Most tools assume a human face. If your character is stylized or non-human, facial animation, lipsync, and expression often fail—or don’t happen at all. What the tools can (and can’t) do: — Higgsfield & MultiTalk ✅ Generate talking-heads from still images, even for non-human characters ❌ Locked into preset animation templates—you can’t bring your own motion or framing 📌 That said: I prefer MultiTalk. Stronger output and feels more flexible. In some cases it accomplishes the whole task. — Veo 3 ✅ Great talking performance for non-human characters ❌ Only works in text-to-video ❌ No iteration—every generation is a new variation ❌ Image-to-video mode doesn’t support talking — Runway (Act-One) ✅ Lets you apply facial performance to a pre-animated video ✅ Works with your own camera and body animation ❌ Only functions if the face is human-like enough to be detected The hybrid workflow I’m testing: 1) Take the first frame of the planned video and humanize it using Flux Kontext (10–30%)—just enough for Act-One to detect it as a face. 2) Generate a body/camera motion performance using image-to-video tools. 3) Load the animated video into Runway with Act-One and apply a facial performance to the humanized version. 4) Bring the resulting video into Luma Labs Modify or Runway Retexture and use your original, non-human frame as the reference. I test both Luma and Retexture—whichever gives better results gets used in the final.

7

2 Comments

As the march toward the March 2nd grand opening of the Open Vector public commons project approaches, I want to add to the discourse today explaining a real workflow I am doing right now with #claudecode #mural #firefox and just a pretty vanilla IDE stack. The flow: 1️⃣ I work with Julian (Claude Desktop + Google Drive, Github, and Desktop Commander extensions, and Notion MCP connection) to describe what I want, the behaviors, the intent, the user experience, etc. 2️⃣ Julian makes a PRD for the feature, the whole vision for a discrete, contained interaction and experience model. We riff on it, and I say to turn into phased prompts. 3️⃣ Julian makes the prompts as MD files and puts them in a Prompt Hopper in Notion. I then give them to the agents as appropriate (Siddig does backend and API, Heavy does front end, Sellivan does repo management, etc). 4️⃣ The named agents are set to "dangerously ignore permissions" and I set them to plan mode, and paste the prompt which is usually 300-400 lines. NOTE: I never write the initial prompts, my follow up prompts are 1-2 lines. Julian handles the prompts and continuity. 5️⃣ The agent, i.e. Heavy runs the plan and asks questions as needed, and then we "Accept and bypass permissions" and Heavy YOLOs the code. Why do I trust this? 1) I read the plan, 2) we commit after each prompt, 3) we make a branch per feature, 4) we make a worktree when I want to do something big. 6️⃣ When it's done, I check it in the browser, try it out, and either quickly dictate changes directly to the agent, or I paste the agent's prompt-by-prompt standup back to Julian. In cases like the screenshot, I purposfully went light on the layout spec because I knew I would want to wireframe AFTER getting it in the browser. 7️⃣ We repeat this until I like what I see, adding tweaks and enhancements as we go. And, as this is all being built on the principles and guidelines contained in the CLAUDE.md for each agent, the project VECTOR.md, and the extensive REPO-BIBLE/*.md that dictates architecture, usually any console errors or bugs are immediately resolved. 8️⃣ We do this for an hour or 2, and when the feature is done, we tag the branch, merge it, and start over. 🏁 The above would normally be a 2 week sprint. This is not vibe coding. This is not capricious. This is not unstable. This is the future. I am working on a plan to create cirriculum to scale this to teams of any size. You may be asking: but where do you ideate? Where is the design system, the brand, the visual styles? They are wherever you want them to be, the CSS rendering is abstracted out. Pictured: - Left: Claude Desktop (Julian) giving me its optical assessment of my wireframe out of Mural (right). Julian can literally just look at my Mural sketch and interpret it with 100% accuracy, - Right: a screenshot of my 1st-draft UI screenshotted right out of Firefox and pasted into Mural so I could see it. - Right: A Mural-primitives wireframe of the changes I wanted.

18

3 Comments

New: 10-week course for creatives working with AI tools! I’ve seen firsthand how much trial, error, and late-night experimenting it takes to turn AI tools into real creative projects. The ThinkDiffusion team has spent years helping artists and educators bridge that gap-and now, they’ve packed all that experience into a single, hands-on course. What’s inside? • Live workshops with creative directors, AI educators & industry pros • Learn from Kan Towongsricharoen, Sebastian Kamph, & Matt S. • Build a capstone project (image, video, or sound) with expert guidance • Weekly support sessions and a private cohort community • Sessions with creators using AI in the real world It kicks off June 19! If you want to level up your creative workflow with AI, check out the details below and apply! 👇 https://lnkd.in/dhBuvHK4  #AI #AIArt #AIEducation #ThinkDiffusion #genAI #course #AIworkshops #generativeai #ComfyUI

34

13 Comments

In this weekend read, Ethan Sheung Yan Chiu breaks down the composable rendering, material, particle, and transition systems behind The Monolith Project, a massive 13-scene WebGL experience. Stunning tech and design insights: https://lnkd.in/d8un7qNR

47

TraceTracks is a cross-platform tool that connects the dots between SoundCloud and YouTube - helping you find remixes, reuploads, and covers of tracks with ease. 🔍 Search by track or artist 🎯 Match SoundCloud tracks to YouTube videos ⚡ Real-time view tracking 📊 Built-in analytics + duplicate filtering 🎹 Navigate entire artist discographies with simple key inputs If you ever need to pivot from SoundCloud to YouTube, this tool is a great example of how you can think outside of the box with so many OSINT methods. Try it here: https://lnkd.in/e9YRD33k

12

Luma AI released Ray3 1.4 yesterday. Native 1080p, 4x faster, 3x cheaper than Ray3. Those are the headline numbers, but what matters more is how it performs under pressure. Temporal stability across frames has improved in ways that change how you approach shot sequencing. The modify workflow now extends to 18 seconds with stronger consistency for subjects, objects, and style. Lighting and material rendering show the kind of improvement you notice immediately, not incremental, but tangible. Complex motion sequences generate with far greater predictability. Artifacts that used to kill usable takes appear less frequently. For projects demanding visual continuity across multiple shots, this means fewer wasted iterations and more reliable output. I've been testing it against sequences that previously required multiple passes to achieve coherence. The difference is real. If you're building anything that requires shot-to-shot consistency or precise camera control, this version deserves your attention. https://lnkd.in/gm4D86aX

2

Don't miss this conversation between our CTO Emily Mu and our Director of Product / CLO J. Michael Dockery, in which they explore Retrieval-Augmented Generation (RAG)—what it is, how it works, and why it matters in legal tech. They break down RAG’s core mechanics, its role in tackling context window limits, and why hallucinations in AI models remain a challenge even with retrieval. With real-world examples from legal research and document review, this conversation is a must-watch for anyone at the intersection of law and AI. https://lnkd.in/eNGVYAdG

10